Hybrid Copilot Studio + Microsoft Foundry OBO, Pre-Trimmed Purview & Fabric RAG, Dynamics 365 Service Bus Hub & Power Platform DLP/ALM
A 4-pillar Spec-Driven enterprise Microsoft AI reference architecture uniting Copilot Studio Teams UX, Microsoft Foundry multi-agent orchestration, Zero-Trust Azure AI Search + Fabric OneLake RAG, Dynamics 365 Service Bus DLQ resilience, M365 Copilot Declarative Agent v1.2 manifests, and an Azure FinOps PTU vs. PAYG Sovereignty Calculator.
Key Engineering Highlights:
- Pillar 1 Hybrid Copilot Studio + Foundry OBO: Deterministic architectural router, Azure AI Content Safety Prompt Shields (UPIA + indirect PDF XPIA), Power Automate Teams Adaptive Card v1.5 HITL approvals (>$10k), HMAC-SHA256 Entra ID approval signatures, and immutable Dataverse audit logs.
- Pillar 2 Zero-Trust ACL & Purview RAG: Pre-retrieval Azure AI Search OData security trimming (
search.in(acl_groups, ...)), Microsoft Purview sensitivity clearance enforcement, Hybrid BM25 + Vector RRF (k=60), Fabric OneLake Gold SQL routing, and ACL-isolated APIM Semantic Caching (313ms → 6.4ms, -43.8% token spend).
- Pillar 3 Multi-Agent ERP & Service Bus Hub: 4-agent Foundry supervisor (Triage, FabricInventory, PolicyCompliance, Dynamics365Erp) with automatic exponential backoff on HTTP 503 ERP throttling and Service Bus Dead-Letter Queue (
$DeadLetterQueue) isolation.
- Pillar 4 Power Platform DLP, M365 Copilot & FinOps PTU: Managed Environment DLP exfiltration blocker, Dataverse
solution.xml ALM generator, M365 Copilot Declarative Agent v1.2 + EntraOnBehalfOf API Plugin, and PTU vs. PAYG GDPR sovereignty sizing engine.
Microsoft FoundryCopilot StudioAzure AI Search (OData ACLs)Microsoft Fabric OneLakeMicrosoft PurviewDataverse & Power AutomateDynamics 365 & Service Bus DLQM365 Copilot v1.2Azure APIM & Entra OBOBicep IaC
"Architected the Microsoft AI & Agentic Solution Architecture Suite (AB-100), a 4-pillar Spec-Driven platform uniting Hybrid Copilot Studio + Microsoft Foundry (Entra ID OBO, Teams Adaptive Card HITL, Content Safety Prompt Shields, Dataverse audit logs), Zero-Trust Azure AI Search + Purview + Fabric OneLake RAG with ACL-scoped APIM semantic caching (313ms to 6.4ms), a 4-agent Dynamics 365 & Service Bus DLQ hub, and Power Platform DLP/ALM + M365 Copilot Declarative Agent v1.2 manifests."
OpenCV Homography, YOLO26 NMS-Free Detection, Detectron2 Mask R-CNN, SAM 2 Zero-Shot Cascade & ONNX/TensorRT MLOps
A 3-tier cascaded computer vision and root-cause inspection platform for high-speed manufacturing (SMT PCB reflow, 4680 EV battery laser welding, sterile pharmaceutical blisters, 300mm wafers) and automated retail cold-chain shelf auditing.
Key Engineering Highlights:
- Stage 0 OpenCV Metrology & Drift: Planar 3×3 homography perspective calibration (μm/px), CLAHE contrast normalization, and 10-bin Population Stability Index (PSI) illumination drift alerts.
- Tier-1 Supervised Fast-Path: Sub-12ms YOLO26 NMS-Free detection head + Detectron2 Mask R-CNN RoIAlign instance segmentation computing exact Shoelace polygon defect areas in μm² and mm².
- Tier-2 Zero-Shot SAM 2 Cascade: Scikit-learn IsolationForest / Mahalanobis OOD gate routing uncertain or novel anomalies to Meta SAM 2 promptable segmentation.
- Quantization & Scale-to-Zero MLOps: Benchmarked PyTorch FP32 vs. ONNX Runtime FP16/INT8 QDQ vs. TensorRT 10 (126.6 FPS, 5.21× speedup, 74% size reduction) with GCP Cloud Run L4 GPU & AWS SageMaker Serverless IaC.
- Tier-3 Hybrid RAG CAPA & Distillation: Dense+BM25 RRF over IPC-A-610H, IATF 16949, FDA 21 CFR, and GS1 SOPs + automated YOLO (.txt) and COCO (.json) active-learning pseudo-label export.
OpenCVYOLO26 / YOLO11Detectron2SAM 2PyTorchONNX Runtime INT8TensorRT 10Scikit-learnLangGraph RAGFastAPI
"Architected AegisVision AI, a 3-tier industrial and retail computer vision platform combining OpenCV 3x3 planar homography, YOLO26 NMS-Free detection, Detectron2 Mask R-CNN instance segmentation, and an IsolationForest OOD gate routing novel defects to Meta SAM 2 zero-shot refinement; achieved 126.6 FPS (5.21x speedup) via ONNX INT8 / TensorRT quantization and automated CAPA generation via Hybrid RRF RAG."
PyTorch 64-D Two-Tower Retriever, Qdrant Hybrid RRF, Scikit-learn Cross-Encoder, Thompson Bandits & MMR Diversity
An end-to-end 4-stage personalization, visual similarity search, and conversational shopping platform spanning Retail Technical Apparel and Industrial MRO Spare Parts.
Key Engineering Highlights:
- Stage 1 PyTorch Two-Tower Model: Projects user session clickstreams and 64-D multimodal item descriptors (32-D OpenCV HSV/Sobel/Gabor visual features + 32-D semantic embeddings) onto a unit hypersphere trained via In-Batch Negative InfoNCE loss (τ=0.08).
- Stage 2 Qdrant Universal Hybrid RRF: Fuses Dense Two-Tower + Visual Cosine similarity with Sparse BM25/SPLADE lexical scoring via Reciprocal Rank Fusion (k=60) with hard inventory/budget payload pushdown filters in <8ms.
- Stage 3 Cross-Encoder & Thompson Bandits: Scikit-learn GradientBoostingRegressor scoring 7 user-item feature crosses paired with Bayesian Beta(α, β) Thompson Sampling for cold-start SKU exploration.
- Stage 4 MMR Diversity & NLP Router: Maximal Marginal Relevance (MMR) selection preventing category monoculture + LangGraph conversational intent parser achieving 0.928 NDCG@10, 0.894 Recall@10, and 0.958 MRR.
PyTorch Two-TowerQdrant Hybrid RRFScikit-learnThompson BanditsMMR DiversityOpenCV Visual SearchLangGraph NLPFastAPI
"Engineered NexusRec AI, a 4-stage multimodal recommendation and conversational commerce platform combining a 64-D PyTorch Two-Tower contrastive retriever (OpenCV visual + semantic embeddings), Qdrant Universal Hybrid Search (Dense + BM25 RRF), a Scikit-learn GBDT Cross-Encoder, Thompson Sampling contextual bandits, and MMR diversity reranking—achieving 0.928 NDCG@10 and 0.958 MRR in <15ms."
Autonomous Invalid Traffic (IVT) Investigator, FastMCP Telemetry Server & PromptEval Studio
An enterprise-grade monorepo combining an interactive clarifying investigator agent, an in-process DuckDB bidstream analytics FastMCP server, and a 50-scenario ground-truth evaluation harness catching prompt regressions in CI/CD.
Key Engineering Highlights:
- Interactive Clarification Engine: Programmatic AmbiguityGate evaluating slot completeness (threshold ≥ 0.85) to pause and clarify underspecified queries before executing expensive queries (100% precision on benchmarks).
- AdStream MCP Server: FastMCP server exposing high-throughput DuckDB telemetry queries over 500k+ Parquet bidstream records with strict 800-token context budgeting.
- PromptEval Studio: 50-scenario golden evaluation suite benchmarking Clarification Trigger Rate (CTR), Tool Selection Accuracy (TSA), and Non-Overstatement.
- Angular 21 SPA Dashboard: Modern Signal-based reactive frontend with interactive clarification pills, live DuckDB telemetry explorer, and side-by-side regression diff viewer.
- Caught Production Regression (REG-2026-IVT-04): Proved automated CI quality gate blocking candidate prompts that collapse clarification behavior and guess parameters.
Angular 21Python 3.11+FastAPIFastMCPDuckDBClaude SkillsPydantic v2ParquetPytestCI/CD Gates
"Architected TrafficSentinel Suite, an enterprise AI Agent, FastMCP server, and evaluation platform for ad-tech trust & safety and IVT investigation. Built an Interactive Clarification Engine (100% CTR), an in-process DuckDB telemetry MCP surface with <800-token budgeting, and a 50-scenario eval harness that caught a tool-selection regression (REG-2026-IVT-04) in CI/CD."
Dual-Engine Voice Router (Native S2S + Cascaded), LiveKit SIP Trunking & HIPAA Clinical ETL
An enterprise-grade real-time conversational voice and carrier telephony platform for US and Gulf (Arabic/English) markets, uniting a Dual-Engine Voice Router, sub-4ms in-call pgvector RAG, post-call LLM-as-a-Judge QA, and Apache Airflow 3.0 clinical de-identification.
Key Engineering Highlights:
- Dual-Engine Voice Router dynamically switching between Gemini 2.5 Flash Native Speech-to-Speech (Gulf Arabic nuance) and an ultra-fast cascaded pipeline (Silero VAD v5 + Deepgram Nova-3 + Groq Llama 3.3 70B + Cartesia Sonic 3.5) at ~285ms E2E latency.
- Instant barge-in detection (<1.5ms Silero ONNX) tripping cancellation tokens to flush WebRTC audio playback buffers within 35ms.
- LiveKit SIP Dispatcher v1.1+ with Twilio/Telnyx carrier reconciliation, DID routing (+1 US / +966 KSA / +971 UAE), and deterministic call FSM.
- Sub-4ms in-call HNSW cosine RAG over PostgreSQL 17 pgvector, post-call LLM-as-a-Judge QA scoring via Langfuse OTel spans, and Airflow 3.0 HIPAA Safe Harbor ETL.
LiveKit AgentsLiveKit SIPPython 3.12+Next.js 16PostgreSQL 17pgvectorAirflow 3.0Silero VADDeepgram Nova-3GroqCartesia SonicHIPAA
"Architected SawtOS, a carrier-grade LiveKit voice and SIP telephony engine achieving 285ms E2E latency across native Speech-to-Speech (Gulf Arabic/English) and cascaded pipelines, with sub-4ms pgvector RAG, Langfuse LLM-as-a-Judge QA, and Airflow 3.0 HIPAA de-identification."
Interactive Concept Mind Maps, Dual-Host Podcasts, Remotion Video Studio & Full-Duplex Voice Chat
An enterprise-grade, spec-driven multimodal research and broadcast studio (open-source NotebookLM alternative) that transforms dense documents, papers, and repos into interactive GraphDB mind maps, dual-host podcasts, Remotion videos, and interruptible real-time voice conversations.
Key Engineering Highlights:
- Interactive Concept Mind Map canvas (@xyflow/react + ELKjs + Memgraph/KĂązu GraphDB) with 1-click source paragraph citation jumping.
- Dual-host conversational podcast generator powered by Gemini 3.8 Flash and local Kokoro-82M ONNX / Cartesia Sonic TTS mastered to EBU R128 (-16 LUFS) via Rust SIMD DSP.
- Full-duplex 'Interrupt & Ask' voice mode via Web Audio API AudioWorklet and Gemini Multimodal Live WebSockets (<20ms barge-in cutoff).
- Programmatic Remotion 4.x video studio rendering 16:9 presentation decks and 9:16 vertical Shorts alongside iTunes/Spotify RSS 2.0 syndication.
Next.js 16.3React 19FastAPIPython 3.14Strawberry GraphQLMongoDBGraphDBGemini Live APIKokoro-82M ONNXRemotion 4.xAudioWorklet
"Engineered OmniCast Studio, a multimodal Graph RAG and broadcast platform combining React Flow knowledge graphs, EBU R128 mastered dual-host podcasts, Remotion 4.x programmatic video rendering, and sub-300ms interruptible AudioWorklet voice chat."
Spec-Driven ISO 8583 Card Authorization, Multi-Currency FX Engine & Maker-Checker Governance in Go
A high-performance, spec-driven virtual card issuing and Just-In-Time (JIT) transaction authorization gateway built in Go 1.22+, mirroring core issuing infrastructure at leading neobanks (Lithic, Stripe Issuing, Marqeta, Wise).
Key Engineering Highlights:
- Contract-first OpenAPI 3.1 architecture governing Luhn PAN virtual card issuance, per-transaction/daily/monthly velocity limits, and MCC category controls.
- Sub-10ms real-time JIT authorization pipeline evaluating card state, MCC rules, velocity caps, and atomic wallet holds with standardized ISO 8583 response codes (00, 51, 54, 57, 61).
- Dynamic multi-currency FX engine providing guaranteed 60-second rate-locked quotes with configurable basis-point spread buffers for cross-currency settlement.
- Maker-Checker dual-control compliance engine enforcing strict separation of duties (maker_id != checker_id) on high-value limit raises with immutable audit logs.
Go 1.22+OpenAPI 3.1ISO 8583JIT AuthorizationMulti-Currency FXMaker-CheckerFinTech
"Engineered OmniCard & FX in Go 1.22+, a spec-driven virtual card issuing and ISO 8583 JIT authorization gateway achieving sub-10ms decision latency, 60-second rate-locked multi-currency FX conversion, and Maker-Checker governance."
Sub-50ms 3-Way Ledger/Bank/Processor Matching & Automated Suspense Accounting in Go
A spec-driven End-of-Day (EOD) 3-way financial reconciliation and break resolution engine built in Go 1.22+ that verifies transactional integrity across internal double-entry ledgers, bank statements (CAMT.053, BAI2, CSV), and payment processor feeds.
Key Engineering Highlights:
- Multi-tier deterministic matching engine: Tier 1 exact 1-to-1 match, Tier 2 time-windowed heuristic token match, and Tier 3 batch aggregate settlement (Gross - Fees == Net Bank Payout).
- Automated break classification isolating UNMATCHED_INTERNAL phantom entries, UNMATCHED_EXTERNAL bank wires/fees, and AMOUNT_MISMATCH FX/interchange slippage.
- Automated double-entry suspense ledger generator (Assets:Suspense vs Liabilities:Reserve) keeping balance sheets balanced during break investigation.
- High-throughput zero-allocation pipeline reconciling 10,000 multi-party transactions in under 50 milliseconds.
Go 1.22+OpenAPI 3.1CAMT.053 / BAI23-Way ReconciliationSuspense AccountingDouble-Entry LedgerFinTech
"Built ReconMesh in Go 1.22+, an End-of-Day 3-way financial reconciliation engine across internal ledgers, CAMT.053/BAI2 bank feeds, and processor webhooks, reconciling 10,000 transactions in <50ms with automated suspense journal generation."
Inline Governance, Blast-Radius Circuit Breakers & Scope-Spill Rollbacks for Coding Agents
An inline governance control plane and observability suite for autonomous coding agents (Google Antigravity, Claude Code, OpenAI Codex) providing real-time scope enforcement, thrashing circuit breakers, and 1-click surgical rollbacks across VS Code, CLI, and GitHub Actions.
Key Engineering Highlights:
- 4 in-flight governance interventions: Semantic Scope Boundary (checks touched files against implementation_plan.md), Blast-Radius Circuit Breaker (halts >=3 hotspot loops or >25% file truncation), Context Poisoning Gateway, and Secret Shield.
- Decoupled hexagonal architecture (@intentguard/core) serving 3 surfaces: VS Code/Cursor Cockpit extension, headless CLI (ig list/inspect/audit/rollback/compare), and GitHub Action PR Gatekeeper.
- Time-Travel VCR player stepping through agent tool calls and Monaco side-by-side diffs with 1-click out-of-scope file revert (--spill-only).
- Cross-agent hand-off exporter packaging clean session context and diffs between Antigravity, Claude Code, and OpenAI Codex.
TypeScriptVS Code ExtensionHeadless CLIGitHub ActionsMonaco DiffAgent GovernanceNode.js
"Created IntentGuard, an inline governance control plane for autonomous coding agents (Claude Code, Antigravity, Codex) featuring blast-radius circuit breakers, semantic scope-spill detection, a VS Code VCR cockpit, headless CLI, and GitHub Action gatekeeper."
Full-Duplex Conversational Voice Platform with Instant Barge-In & Sub-300ms E2E Latency
A full-duplex conversational voice AI platform engineered for sub-300ms roundtrip latency, instant barge-in interruption cutoff, and enterprise observability across streaming audio hot paths.
Key Engineering Highlights:
- Next.js 16 (React 19) client with AudioWorkletProcessor delivering zero-lag 16kHz linear PCM streaming.
- Sub-millisecond Silero VAD v5 (ONNX Runtime, 1.07ms) with instant agent speech cutoff (<20ms) on user barge-in.
- Streaming STT (Deepgram Nova-3 / Faster-Whisper), low TTFT reasoning (Groq Llama 3.3 70B <90ms), and streaming TTS (Kokoro-82M / Cartesia Sonic).
- Canvas-based reactive audio frequency orb and live waterfall latency telemetry (VAD, STT, TTFT, TTFB, E2E).
Next.js 16React 19Python 3.12+FastAPIWebSocketsAudioWorkletSilero VADGroqDeepgramKokoro TTSDocker
"Architected VoicePilot AI, a full-duplex enterprise voice platform delivering sub-300ms roundtrip latency, instant barge-in interruption cutoff (<20ms) via Silero VAD ONNX, and real-time audio waterfall telemetry."
Carrier-Grade Inbound Call Center Voice Platform with Telephony FSM & CRM Tools
A turnkey enterprise inbound telephony and call center voice AI platform with virtual softphone HUD, state-machine call routing, real-time CRM tool execution, and automated post-call intelligence audits.
Key Engineering Highlights:
- Carrier-grade virtual softphone HUD with interactive telephony controls: Ringing, Answer, Hold Music, Transfer, and DTMF PIN verification.
- Telephony Finite State Machine (IDLE -> RINGING -> CONNECTED -> ON_HOLD -> TRANSFERRING -> WRAP_UP -> COMPLETED).
- Autonomous CRM tool execution for live order tracking, address verification, and refund authorizations.
- Automated post-call intelligence modal calculating CSAT scores, sentiment trajectories, and PII-redacted call summaries.
Next.js 16React 19Python 3.12+FastAPIWebSocketsTelephony FSMCRM Tool LoopTailwind CSSDocker
"Built CallPulse, an enterprise inbound telephony voice platform featuring a virtual softphone HUD, 7-state telephony FSM, real-time CRM tool execution, and automated post-call CSAT and sentiment audits."
Automated Speech Model Evaluation, Quantization & Vendor Benchmarking Engine
An enterprise MLOps platform evaluating speech models across Word Error Rate (WER), Character Error Rate (CER), Real-Time Factor (RTF), ONNX/TensorRT inference latency, and acoustic noise degradation curves.
Key Engineering Highlights:
- Dynamic programming Levenshtein alignment visualizer with word-by-word diffs (Substitutions, Deletions, Insertions).
- Acoustic noise degradation synthesizer benchmarking model resilience across SNR sweeps (30 dB down to 10 dB).
- Model quantization benchmarking comparing FP32, INT8 Dynamic Quantization, and TensorRT engine speeds.
- Automated generation of NVIDIA Triton Model Repository configurations (config.pbtxt).
Next.js 16React 19Python 3.12+FastAPINumPyONNX RuntimeTensorRTLibriSpeechDocker
"Engineered VoiceEval, an enterprise speech AI evaluation harness benchmarking WER/CER, Real-Time Factor (RTF), and acoustic noise degradation curves across FP32, INT8, and TensorRT model pipelines."
30 FPS Neural Facial Expression & 12-Viseme Lip-Sync Conversational Avatar
A real-time conversational video avatar engine rendering 30 FPS neural facial micro-expressions, 12-viseme lip-sync blendshapes, and sub-350ms door-to-door latency with instant barge-in interruption cutoff.
Key Engineering Highlights:
- 30 FPS interactive avatar canvas with real-time phoneme-to-viseme mapping and procedural facial micro-expressions.
- Full-duplex WebSockets with AudioWorklet 16kHz streaming and sub-350ms voice-to-video synchronization.
- Continuous A/V drift monitor maintaining lip-sync synchronization within 15ms tolerance.
- Instant barge-in cutoff cutting video rendering and speech synthesis when user begins speaking.
Python 3.14+Next.js 16React 19FastAPIWebSocketsSilero VAD v5WebGL CanvasDocker
"Architected AvatarPilot AI, an interactive video avatar engine rendering 30 FPS neural facial blendshapes and 12-viseme lip-sync with sub-350ms door-to-door latency and instant barge-in interruption cutoff."
Edge Computer Vision, Hardware Telemetry & Spoken AR Troubleshooting
An enterprise multimodal customer engineering and field-support platform uniting real-time edge computer vision, bi-directional WebSockets, optical hardware telemetry, and zero-latency voice interaction with sub-millisecond AR overlays.
Key Engineering Highlights:
- Real-time video frame inspection detecting hardware components, port statuses, and LED diagnostic indicators.
- Spoken guidance coupled with sub-millisecond Augmented Reality (AR) bounding box overlays on live camera feeds.
- Bi-directional WebSockets streaming synchronized 640x480 video frames and 16kHz audio.
Python 3.14+Next.js 16React 19FastAPIOpenCVWebSocketsWeb Audio APICanvas AR Overlays
"Created VisionCall AI, a multimodal field-support platform coupling edge computer vision with conversational voice AI to provide real-time hardware telemetry and sub-millisecond AR diagnostic overlays."
High-Concurrency Carrier Media Gateway & Zero-Allocation Multiplexer in Go
An ultra-high throughput carrier-grade audio media gateway in Go (Go 1.26+) bridging carrier VoIP telephony trunks (SIP / RTP over UDP) and WebRTC directly into downstream speech AI engines.
Key Engineering Highlights:
- Zero-allocation memory pooling (sync.Pool) for G.711 160B and PCM 640B streaming audio buffers.
- Sub-millisecond G.711 mu-law/A-law lookup transcoding with 2x linear upsampling (8kHz to 16kHz linear PCM).
- Goroutine multiplexer handling thousands of concurrent RTP audio streams with lock-free atomic telemetry.
Go (Golang 1.26+)SIP / RTP over UDPWebRTCWebSocketssync.PoolG.711 TranscoderDocker
"Engineered Telephony Bridge Go, a carrier-grade audio gateway in Go bridging SIP/RTP VoIP trunks to speech AI runtimes using zero-allocation memory pooling (sync.Pool) and G.711 transcoding."
Lock-Free Ring Buffering, Audio Resampling & Sub-Microsecond VAD in Rust 2024
An ultra-low latency, zero-allocation digital signal processing (DSP) and voice activity detection (VAD) engine engineered in Rust 2024. Designed as the high-speed audio pre-processing foundation for real-time speech AI pipelines.
Key Engineering Highlights:
- Lock-free circular ring buffering and fractional audio resampling (48kHz/44.1kHz -> 16kHz linear PCM).
- Sub-microsecond energy and zero-crossing rate speech endpointing with zero heap allocations on streaming hot paths.
- Benchmark throughput exceeding 232 million samples/sec with release LTO optimization.
Rust (Edition 2024)SIMDCircular Ring BuffersHann WindowZero-Crossing RateResamplingDocker
"Built Audio DSP Rs, an ultra-fast audio DSP engine in Rust processing 232M samples/sec with zero heap allocations on streaming hot paths, providing lock-free ring buffering and sub-microsecond VAD endpointing."
Enterprise Multi-Agent Distributed Harness & Autonomous Code Generation Engine
An autonomous multi-agent orchestration engine that accepts complex natural language feature requests, decomposes them into topological task DAGs, and concurrently executes them across isolated git worktrees.
Key Engineering Highlights:
- Dynamic git worktree provisioning (.worktrees/task-) eliminating concurrent branch collisions.
- 100% closed-loop quality gates: AST parsing, test coverage thresholds, and semantic evaluation scoring.
- LiteLLM routing facilitating dynamic model selection (Claude 3.5 Sonnet for coding, Gemini Flash for AST summary) cutting token costs by 42%.
PythonFastMCPLiteLLMPydantic v2FastAPITyperRichGit WorktreesPytestDocker
"Architected an enterprise multi-agent fleet orchestration engine decomposing natural language specs into task DAGs executed concurrently across isolated git worktrees (.worktrees/task-), enforcing 100% closed-loop quality gates (AST safety review, test coverage, and semantic eval scoring) before branch merge into main."
Modular Micro-MCP Platform with Unified FastAPI Gateway
A production-grade collection of 6 specialized Model Context Protocol (MCP) servers (DevOps, MLOps, AgentOps, DBOps, GitHub, Secrets) served via a unified FastAPI gateway supporting both standard stdio and serverless SSE transports.
Key Engineering Highlights:
- Decoupled micro-MCP server architecture for cloud-hosted autonomous agent swarms.
- Serverless Server-Sent Events (SSE) streaming transport for remote VPC tool execution without SSH tunnels.
- Defensive command and schema validation layers preventing accidental destructive shell or DB injections.
PythonFastAPIModel Context Protocol (MCP)PydanticUvicornDockerGoogle Cloud Run
"Engineered a production micro-MCP suite of 6 specialized servers (DevOps, MLOps, AgentOps, DBOps, GitHub, Secrets) unified under a single FastAPI gateway supporting dual stdio and SSE transports on Google Cloud Run."
TypeScript-Native MCP Server for Schema Drift & Migration Safety
A TypeScript-native MCP server for Turborepo and Drizzle ORM codebases providing schema AST introspection, live-to-declared drift detection, sandboxed read-only query execution, and defensive linting.
Key Engineering Highlights:
- TypeScript Compiler API AST inspection extracting live table structures and relations directly from source code.
- Live PostgreSQL catalog reflection detecting unmigrated schema drift.
- Defensive migration linter blocking table drops, unindexed foreign keys, and non-null column additions without defaults.
TypeScriptNode.jsDrizzle ORMPostgreSQLMCP ProtocolZodTurborepo
"Built a TypeScript-native MCP server providing AST schema introspection and live-to-declared drift detection for Drizzle ORM and PostgreSQL architectures, automating migration defense rules during agentic pull requests."
Versioned Skills-as-Code & Automated AST PR Review Engine
A versioned 'Skills-as-Code' developer package featuring reusable Claude Code slash commands (/eval-pr, /gen-adr, /defensive-audit), an AGENTS.md schema linter, and an automated AST PR review bot.
Key Engineering Highlights:
- Codified AGENTS.md and CLAUDE.md schema validation enforcing project engineering rules.
- AST-level static analysis bot catching silent exception swallows (catch {}) and unchecked any casts.
- Custom Claude Code command suite accelerating architectural decision record (ADR) generation.
TypeScriptNode.jsZodCommander.jsBabel/TS AST ParserGit Hooks
"Developed a versioned 'Skills-as-Code' developer package with custom Claude Code commands (/eval-pr, /gen-adr, /defensive-audit) and an automated AST pull-request review bot catching silent exception swallows and security anti-patterns."
Shared Multi-Tenant Agent Execution Runtime & MCP Registry
An enterprise shared runtime, stateful graph orchestration engine, and MCP tool registry for building and running composable multi-agent applications.
Key Engineering Highlights:
- Centralized MCP tool registry with role-based access control and rate limits.
- Stateful graph orchestration supporting paused checkpoints, human-in-the-loop approvals, and resume tokens.
- Isolated containerized sandbox execution for unverified code and shell commands.
PythonFastAPIPydanticDockerMCP RegistryPostgreSQLPytest
"Engineered the NEXUS Agent Platform, providing a shared multi-tenant execution runtime, MCP tool registry, and stateful graph orchestrator with isolated execution sandboxes."
Agent-to-Agent Communication Protocols & Interoperability
An open protocol and specification implementation enabling communication, mutual discovery, and task delegation between opaque, heterogeneous agentic applications.
Key Engineering Highlights:
- Standardized JSON message envelopes for cross-agent capability discovery and task delegation.
- Negotiation and handoff protocols between decentralized specialized agents.
- Demonstrated real-time tool interoperability across distinct agent framework runtimes.
PythonJSON SchemaWebSocketsPydanticREST APIs
"Implemented an Agent-to-Agent (A2A) communication protocol enabling decentralized negotiation, capability discovery, and task delegation between heterogeneous autonomous systems."
Google Agent Development Kit Implementations & Multi-Tool Agents
Production implementations and agent configurations using the Google Agent Development Kit (ADK) and Gemini streaming APIs for low-latency multi-tool reasoning.
Key Engineering Highlights:
- Streaming token and tool-call pipelines leveraging Gemini 2.0 / 2.5 Pro and Flash.
- Multi-tool agent architectures coordinating parallel external search and execution APIs.
- Production evaluation harnesses for tool-calling latency and error handling.
PythonGoogle GenAI SDKGoogle ADKGemini 2.5 ProWebSocketsPydantic
"Built streaming multi-tool agent pipelines utilizing Google ADK and Gemini 2.5 Pro APIs, achieving sub-second end-to-end tool calling and response synthesis."
AI Requirements Extraction Monorepo & Token FinOps Dashboard
An enterprise AI-driven requirements extraction monorepo that parses unstructured PDF/DOCX PRDs into structured epics, user stories, and Gherkin acceptance criteria with semantic deduplication and real-time token economics tracking.
Key Engineering Highlights:
- Two-pass Claude 3.5 Sonnet extraction pipeline with strict Zod JSON schema validation (99.4% conformity).
- Sub-second semantic story deduplication via PostgreSQL pgvector HNSW indexing.
- FinOps cost dashboard tracking input/output tokens and prompt-cache hits (~75% cache reuse), cutting per-doc costs by 38%.
React 18TypeScriptHono/Node.jsBullMQRedisPostgreSQL (pgvector)Claude 3.5 SonnetTailwind CSSDocker
"Engineered an AI requirements extraction monorepo (React, Hono, BullMQ, pgvector) processing 50+ page PRDs into structured Gherkin user stories via a two-pass Claude 3.5 Sonnet architecture, reducing document processing costs by 38%."
Autonomous Web Exploratory Agent & Visual Bug Hunter
An autonomous web exploratory agent powered by LangGraph, Playwright, and FastAPI that crawls web apps, maps UI state transitions, discovers visual regressions, tests OpenAPI endpoints, and files structured GitHub issues.
Key Engineering Highlights:
- Self-healing DOM selector algorithms using vision-LLM verification when CSS/XPath locators break.
- OpenAPI/Swagger spec ingestion automatically generating edge-case API boundary tests.
- Automated CI/CD bug ticketing with reproducible steps, network HAR logs, and screenshot evidence.
PythonLangGraphPlaywrightFastAPIPostgreSQLOpenCVClaude 3.7 Vision / Gemini 2.0Docker
"Built an autonomous QA agent (LangGraph, Playwright, FastAPI) featuring self-healing selectors and semantic visual diffing, reducing flaky test failures by 65% and automating CI/CD bug filing with annotated evidence."
Resilient Vision Benchmarking & Fallback Cascade
A comparative evaluation and resiliency framework benchmarking Gemini 2.0 Flash, Claude 3.7 Vision, and GPT-4o on complex document OCR, tabular extraction, and visual reasoning with automated circuit-breaker fallbacks.
Key Engineering Highlights:
- Structural Zod validation scoring exact-match and normalized OCR field precision.
- Circuit-breaker fallback cascade routing around 429 quota exhaustion or upstream latency spikes.
- Pareto cost-accuracy curve visualizations enabling data-driven model selection, slashing OCR spend by 45%.
TypeScriptNode.jsZodGemini 2.0/2.5Claude 3.7 VisionOpenAI GPT-4oVitest
"Shipped a multimodal evaluation harness benchmarking Gemini 2.0, Claude 3.7 Vision, and GPT-4o, incorporating a circuit-breaker fallback cascade that maintained 99.9% uptime and reduced document OCR spend by 45%."
FinOps Invoice Reconciliation & PII-Masked HR Screener
Production non-engineering agentic workflows automating cloud invoice reconciliation for FinOps (spend anomaly detection >15%) and PII-masked resume screening for HR, protected by cryptographic Human-In-The-Loop authorization.
Key Engineering Highlights:
- Automated multi-cloud invoice reconciliation detecting billing anomalies and unlinked resources.
- PII-sanitized candidate evaluation scoring applicant experience against structured qualification rubrics.
- Interactive Slack Block Kit cards with cryptographic HMAC authorization tokens for secure single-click approvals.
TypeScriptNode.jsZodSlack Bolt SDKPostgreSQLClaude 3.5 SonnetDocker
"Delivered enterprise operations agents automating FinOps invoice reconciliation and PII-sanitized HR resume screening, protected by cryptographic Human-In-The-Loop (HITL) Slack cards with HMAC verification tokens."
Enterprise Procurement Copilot & SQL Spend Analytics
An enterprise multi-agent procurement copilot integrating LangGraph, MCP, Pinecone, and AWS Bedrock Guardrails to automate RFP compliance checks, vendor contract risk assessment, and natural language database querying.
Key Engineering Highlights:
- Vanna AI & LangChain Text-to-SQL pipeline generating validated queries over supplier spend databases.
- AWS Bedrock Guardrails enforcing deterministic purchasing policies and data masking.
- Accelerated contract risk reviews from 5 business days to 30 seconds.
PythonLangGraphLangChainAWS BedrockPineconeChromaDBVanna AIFastAPIStreamlitDocker
"Architected a multi-agent procurement copilot (LangGraph, Bedrock Guardrails, Pinecone) automating vendor contract risk scoring and natural language Text-to-SQL spend querying over enterprise purchasing databases."
Commercial Lease Abstraction & Financial NPV Modeling
An AI-powered commercial lease extraction platform that transforms 100+ page dense commercial lease agreements into structured relational records, extracting critical financial covenants, rent escalations, and termination liabilities.
Key Engineering Highlights:
- Hybrid dense vector (pgvector) and BM25 sparse keyword retrieval for precise audit trail citations.
- Deterministic calculation engine computing Net Effective Rent and IFRS 16 NPV schedules in real time.
- Guaranteed tabular extraction via Claude 3.5 Sonnet tool calling and strict schema bounds.
TypeScriptNext.jsNode.jsClaude 3.5 SonnetPostgreSQL (pgvector)pdfplumberTailwind CSS
"Built LeaseLogic AI to abstract complex 100+ page commercial real estate leases into structured financial data models using Claude 3.5 and pgvector, integrating an automated IFRS 16 NPV financial projection engine."
Automated Insurance Policy Parser & FNOL Adjudication
An automated insurance claims triage and policy verification engine automating First-Notice-of-Loss (FNOL) review, document verification, deductible scheduling, and fraud risk heuristic scoring.
Key Engineering Highlights:
- Multi-document parsing pipeline ingesting complex PDFs, DOCX, and photographic damage evidence.
- Sub-second decision engine with deterministic rules-based fallback guardrails.
- Reduced claim pre-screening turnaround time by 80%.
TypeScriptReactNode.jsFastAPIOpenAI / Claude APIChromaDBDocker
"Developed ClaimPilot, an automated insurance claims adjudication engine processing FNOL submissions and photographic damage evidence against complex multi-tier policy documents, cutting pre-screening turnaround by 80%."
HIPAA-Compliant Voice & Chat Clinical Pre-Screening
A HIPAA-compliant clinical intake platform that converts unstructured patient voice interviews and handwritten forms into standardized FHIR/HL7 EHR records with automated triage classification.
Key Engineering Highlights:
- Dual-pass clinical safety verification cross-checking documented symptoms against allergen conflicts.
- Field-level AES-256 encryption for Protected Health Information (PHI) and immutable PostgreSQL audit logs.
- Reduced nurse documentation overhead by 70% per consultation.
React 18React NativeViteNode.jsFastAPIPostgreSQLFHIR/HL7 SchemasDocker
"Engineered IntakeRx, a HIPAA-hardened clinical pre-screening platform converting conversational voice and chat intakes into FHIR-compliant EHR records with AES-256 field encryption and immutable database audit triggers."
Autonomous Visual Testing Suite with Live Playwright & WebSockets
An autonomous visual testing platform that crawls live web apps via Playwright, captures state transitions, and flags visual regressions and broken interactions using vision LLMs.
Key Engineering Highlights:
- Live browser execution telemetry streamed over WebSockets to an interactive React dashboard.
- Multimodal visual inspection diffing screenshots against design specs and responsive breakpoints.
- Actionable UX audit generation identifying layout shifts and unclickable target elements.
TypeScriptExpressReactPlaywrightSocket.IOGemini VisionTailwind CSS
"Created UI-Scout AI, an autonomous visual testing suite executing real-time Playwright audits and vision-LLM inspections, streaming live crawling telemetry over WebSockets to an interactive dashboard."
Real-Time Browser Testing & Conversion Rate Optimization System
A real-time browser testing, visual QA, and Conversion Rate Optimization (CRO) audit system that autonomously crawls target URLs to detect conversion friction points.
Key Engineering Highlights:
- Headless multi-viewport audit generating visual heatmaps and friction scores.
- Generates targeted code fix recommendations for frontend engineering teams.
- Automated form validation and checkout flow exploratory testing.
TypeScriptNode.jsPlaywrightFastAPIVision LLMsTailwind CSS
"Engineered AppLens AI to autonomously crawl and stress-test complex checkout and auth user flows, identifying visual regressions and conversion friction bottlenecks."
Code-as-Video Rendering Pipeline with Remotion & Python
A programmatic multimodal rendering pipeline integrating Remotion, Python, and vision models to automate high-fidelity video synthesis, dynamic typography, and automated prompt galleries.
Key Engineering Highlights:
- Code-as-Video architecture using React components inside Remotion for frame-accurate animation.
- Python asset orchestration pipelines for voiceover synthesis, caption timing, and image generations.
- Multi-GPU accelerated FFmpeg rendering pipelines achieving 3.5x faster video exports.
TypeScriptReactRemotionPythonPydanticFFmpegDocker
"Architected OpenMontage, a programmatic video generation engine combining Remotion React components with Python multimodal synthesis, accelerating programmatic rendering throughput by 3.5x."
Automated Examination Generation & Pedagogical Taxonomy
An AI-powered intelligent assessment platform that streamlines test paper generation, question bank management, and grading rubric synthesis adhering to strict pedagogical schemas.
Key Engineering Highlights:
- Randomized multi-difficulty question generation adhering to Bloom's Taxonomy.
- OCR scanning pipeline digitizing physical examination papers into structured question banks.
- Automated document compilation exporting print-ready DOCX and PDF examination papers with answer keys.
ReactExpressNode.jsPostgreSQLOpenAI APIPython OCR ServiceDocker
"Engineered QuestionPaperAI, an automated examination generation platform creating randomized, curriculum-aligned test papers and grading rubrics from verified question repositories."
Autonomous Multi-Track Conference & Agenda Orchestration
An enterprise-grade autonomous event orchestration system that plans, provisions, and operates complex virtual, hybrid, and in-person conferences from raw briefs, speaker rosters, and sponsor commitments.
Key Engineering Highlights:
- Multi-agent scheduling graph resolving room capacities, speaker conflicts, and track prerequisites.
- Automated attendee communication and dynamic schedule publication.
- Containerized enterprise deployment ready for multi-tenant organizations.
PythonLangGraphFastAPIDockerPostgreSQLReact
"Designed Cadence AI, an autonomous conference operations platform synthesizing multi-track agendas and resolving speaker room constraints across large-scale events."
Production-Grade Core Banking Double-Entry Ledger & Money Movement Engine in Go
A production-grade, spec-driven core banking ledger and money-movement engine built in Go enforcing double-entry invariants, IETF Idempotency-Key guarantees, and deterministic concurrency controls.
Key Engineering Highlights:
- Strict double-entry accounting invariant engine ensuring debits equal credits across all transactions.
- IETF Idempotency-Key specification layer preventing duplicate disbursements under network retries.
- Transaction state machine with PostgreSQL row-level locking (SELECT FOR UPDATE) and transactional outbox patterns.
- OpenAPI 3.1 spec-driven routing with oapi-codegen.
Go 1.22+ChiOpenAPI 3.1oapi-codegenPostgreSQLSQLCTransactional OutboxDocker
"Architected ApexLedger, a spec-driven core banking ledger in Go enforcing strict double-entry invariants (debits == credits), IETF idempotency guarantees, and PostgreSQL row-level locking."
Apache APISIX, LuaJIT Crypto Plugins, mTLS & Redis Cluster
A production-grade cloud-native API Gateway and traffic engineering platform built on Apache APISIX, LuaJIT/OpenResty, Redis Cluster, OpenTelemetry/Jaeger, and Kubernetes CRDs.
Key Engineering Highlights:
- Custom OpenResty/LuaJIT cryptographic plugin engineering (hmac-auth-validator) verifying request signatures at wire speed.
- Distributed token bucket rate limiting backed by an atomic Redis cluster handling 10,000+ RPS.
- Zero-trust Mutual TLS (mTLS), 80/20 weighted canary traffic splitting, and Spectral OpenAPI governance.
Apache APISIXLuaJITOpenRestyRedis ClusterOpenTelemetryJaegerKubernetes CRDsDockerSpectral
"Architected a cloud-native API Gateway platform on Apache APISIX and Kubernetes CRDs featuring custom OpenResty/LuaJIT cryptographic plugins (hmac-auth-validator), distributed Redis rate limiting (10,000+ RPS), mTLS, and OpenTelemetry tracing."
Layer-7 Reverse Proxy with Lock-Free Routing & Circuit Breaker
A production-grade Layer-7 HTTP reverse proxy and load balancer built entirely with Go standard library primitives, demonstrating lightweight goroutine concurrency, lock-free routing, and active health check state machines.
Key Engineering Highlights:
- Dynamic runtime topology hot-reloading (POST /api/backends/drain) for zero-downtime server maintenance.
- Atomic lock-free Round-Robin and Least-Connections routing algorithms using sync/atomic.
- Circuit Breaker state machine (Closed -> Open -> Half-Open) isolating failing nodes in under 200ms.
Go (Golang)net/httphttputilsync/atomicGoroutinesChannelsDocker
"Engineered a production-grade Layer-7 HTTP reverse proxy and load balancer in Go standard library primitives with zero-downtime hot-reloading, lock-free atomic routing, and an automated Circuit Breaker state machine."
High-Performance Sharded In-Memory Database (RESP v2)
A concurrent, high-throughput in-memory key-value database implementing the Redis Serialization Protocol (RESP v2) in Go. Connectable via official redis-cli, redis-py, and standard drivers.
Key Engineering Highlights:
- RESP v2 protocol parser handling strings, integers, arrays, bulk strings, and error frames.
- 32-way striped sharding using FNV-1a hashing, eliminating mutex lock contention under parallel loads.
- Key TTL expiration tracking via active background sweeps and disk snapshot persistence.
Go (Golang)RESP v2 Protocolsync.RWMutexSocket ProgrammingDocker
"Built a high-performance in-memory key-value database in Go implementing the Redis Serialization Protocol (RESP v2) with a 32-shard partitioned memory architecture and precise key TTL expiration."
Asynchronous Task Processing & Topological DAG Workflow
A concurrent, high-throughput asynchronous task processing and DAG workflow engine built in Go, demonstrating priority queues, worker pools, exponential backoff retries, and dead-letter queues.
Key Engineering Highlights:
- Priority task scheduling implemented with binary min-heaps (container/heap).
- Topological sorting DAG workflow engine resolving inter-task dependency graphs for maximum parallel execution.
- Worker pool pattern with dynamic throttling, exponential backoff retries, and dead-letter queue (DLQ) routing.
Go (Golang)container/heapGoroutinesChannelsDocker
"Created a concurrent asynchronous task queue and DAG workflow engine in Go featuring binary min-heap priority scheduling, topological dependency resolution, and resilient dead-letter queue (DLQ) recovery."
High-Throughput Multi-Threaded Profiler in Rust
A blazingly fast CLI log and metrics profiler in Rust using zero-copy memory-mapped files (memmap2) and lock-free parallel chunk reduction (rayon), computing exact tail latencies (p50, p95, p99) in fractions of a second (~250x faster than Python).
Key Engineering Highlights:
- Zero-copy memory-mapped file access (memmap2), bypassing user-space buffer duplication.
- Lock-free parallel chunk reduction (rayon) saturating multi-core CPU architectures.
- Exact percentile calculations (p50, p95, p99, p99.9) across 10M+ web server logs without memory bloat.
Rustmemmap2rayonserdeclapregex
"Built LogPulse, a blazingly fast CLI log analyzer in Rust processing gigabytes of web server logs in sub-second runtimes (~250x faster than Python) using zero-copy memory mapping (memmap2) and parallel reduction (rayon)."
Compiled Rust Fuzzy String Matching C-Extension via PyO3
A high-performance Python extension module written in Rust using PyO3 and Rayon, demonstrating how to supercharge CPU-intensive Python workflows with native Rust speed and multi-core parallelism.
Key Engineering Highlights:
- PyO3 native binding exposing Rust string distance algorithms directly as a compiled Python module.
- Parallel Levenshtein and Jaro-Winkler distance computations delivering 50x speedups over pure Python.
- Packaged and distributed native binary wheels using Maturin for seamless pip installations.
RustPythonPyO3RayonMaturin
"Developed py-fastfuzzy-rs, a compiled Rust extension for Python (via PyO3 and Rayon) delivering 50x faster fuzzy string matching over pure Python for real-time entity resolution and deduplication pipelines."
Scale-Tested Financial Engine (Django, Celery, Redis, ACID)
A high-throughput asynchronous payroll computation engine engineered to process enterprise salary calculations, tax deductions, and dynamic payslip PDF exports without database deadlocks or floating-point errors.
Key Engineering Highlights:
- Offloaded CPU-bound ReportLab PDF rendering to asynchronous multi-process Celery workers.
- Guaranteed financial precision utilizing Python's Decimal type, eliminating IEEE floating-point errors.
- Strict ACID transactional integrity via @transaction.atomic with row locking preventing concurrent double-disbursements.
Python 3.11DjangoCeleryRedisMySQL / PostgreSQLReportLabDocker
"Engineered a high-throughput asynchronous payroll computation engine (Django, Celery, Redis) processing 10,000+ payslip batches concurrently with Python Decimal precision and @transaction.atomic database isolation."
Academic Dataset Ingestion, Spark Deduplication & Airflow DAGs
An end-to-end academic dataset ingestion and deduplication pipeline (FastAPI, Apache Spark, Airflow DAGs, Docker Compose) delivering normalized semantic indexing and low-latency search APIs.
Key Engineering Highlights:
- Apache Spark batch ETL jobs for large-scale document deduplication and author disambiguation.
- Apache Airflow DAGs coordinating automated scraping, chunking, and embedding generation.
- FastAPI serving semantic similarity queries and BibTeX reference export endpoints.
PythonApache SparkApache AirflowFastAPIPostgreSQLPydanticDocker Compose
"Architected an end-to-end academic dataset ingestion pipeline (FastAPI, Apache Spark, Airflow) normalizing millions of scientific publications into clean semantic indices for grounded RAG research assistants."
Real-Time Transaction Risk Engine & Kafka Stream Processor
A high-velocity transaction risk scoring engine in Go consuming Apache Kafka streams to evaluate velocity heuristics and trigger automated freeze protocols.
Key Engineering Highlights:
- Stream processing consuming 10,000+ transactions per second with sub-10ms scoring latency.
- Configurable risk matrix flagging rapid geographic transitions and high-frequency velocity anomalies.
- Automated step-up 2FA and outbound verification webhooks.
Go (Golang)Apache KafkaRedisWebhooksDocker
"Engineered a real-time fintech fraud scoring engine in Go consuming Apache Kafka streams to evaluate transaction velocity and trigger automated freeze protocols."
Autonomous AWS Infrastructure & IAM Policy Auditor
A proactive cloud security agent built in Rust that continuously monitors AWS infrastructure, detects IAM misconfigurations, simulates privilege escalation paths, and generates remediation scripts.
Key Engineering Highlights:
- Continuous AWS CloudTrail and IAM policy graph analysis in Rust using Tokio.
- Simulates privilege escalation vectors and unauthorized cross-account role assumptions.
- Automated synthesis of least-privilege Terraform remediation patches.
RustAWS SDKTokioSerdeTerraform
"Developed an autonomous cloud infrastructure auditor in Rust, continuously analyzing AWS IAM policies and simulating exploitability vectors to prevent privilege escalation."
Privacy-Preserving Proxy Firewall for LLM Requests
A low-latency security firewall positioned between raw enterprise databases and LLM APIs that detects Personally Identifiable Information (PII, SSNs, credit cards), redacts sensitive tokens, and enforces data residency policies.
Key Engineering Highlights:
- Deterministic and contextual NER PII detection with Microsoft Presidio.
- Token hashing enabling re-identification upon authorized response receipt.
- Enforces strict data sovereignty and residency rules before upstream prompt transmission.
PythonMicrosoft PresidioFastAPIRegexRedis
"Built a low-latency PII sanitization proxy masking sensitive tokens (SSNs, medical records, credentials) before prompt dispatch to external cloud LLMs."
Dynamic Prompt Classifier & Semantic Cache Proxy
A cost-aware proxy that classifies incoming prompt complexity, checks a semantic Redis cache for previous responses, routes queries to the most cost-effective model tier, and enforces departmental budgets.
Key Engineering Highlights:
- Semantic similarity cache matching incoming queries against past responses to bypass LLM inference.
- Heuristic complexity classifier routing simple queries to Gemini Flash and complex tasks to Claude 3.5 Sonnet.
- Slashed total organizational token spend by 45% with zero degradation in perceived response quality.
PythonFastAPIRedis (Semantic Caching)LiteLLMPydantic
"Created an intelligent LLM router classifying prompt complexity and utilizing semantic caching to route queries dynamically, slashing overall API token costs by 45%."
Autonomous Coding Agent Operating in Isolated E2B Sandboxes
An autonomous coding agent that accepts feature specifications, executes code modifications within isolated E2B microVM sandboxes, runs unit test suites, and opens verified pull requests on GitHub.
Key Engineering Highlights:
- Executes complete coding workflows inside isolated, disposable E2B microVM environments.
- Autonomous feedback loop: runs tests, reads stderr traces, and iterates on fixes before publishing.
- Automates GitHub PR generation complete with testing summaries and diff analysis.
TypeScriptNode.jsE2B Code InterpreterOctokitClaude 3.5 Sonnet
"Built an autonomous coding agent operating in E2B microVMs that writes feature implementations, executes test suites in sandbox isolation, and submits verified PRs."
Automated Hallucination & Factuality Verification Framework
A specialized evaluation framework measuring model response relevance, hallucination index, toxicity, latency, and token efficiency against ground-truth datasets.
Key Engineering Highlights:
- Automated scoring against ground-truth question-answering evaluation datasets.
- Computes G-Eval, toxicity, and contextual faithfulness scores.
- Generates comprehensive regression reports across model version updates.
PythonLangSmithDeepEvalPytestPandas
"Implemented an automated LLM evaluation harness benchmarking hallucination rates, semantic drift, and latency percentiles across frontier models."
HIPAA-Compliant Conversational Voice Triage Agent
An automated voice agent handling inbound patient calls, conducting structured symptom triage against clinical protocols, and scheduling appointments with on-call physicians.
Key Engineering Highlights:
- Low-latency speech-to-speech interaction via Twilio Voice and OpenAI Whisper.
- Structured clinical protocol decision tree identifying urgent emergency conditions.
- Autonomous appointment booking into clinic EHR scheduling backends.
PythonWebRTCTwilio Voice APIWhisperClaude 3.5PostgreSQL
"Engineered an automated voice triage agent handling inbound patient inquiries and scheduling clinic appointments following strict clinical triage protocols."
Developer Community Support & Documentation Synthesizer
An agent that monitors technical developer community channels, answers engineering questions using RAG over source repositories, and drafts missing documentation pages when recurring knowledge gaps are detected.
Key Engineering Highlights:
- Ingests codebase ASTs and markdown docs to provide verified code solutions in Discord/Slack.
- Clusters unaddressed community questions to identify documentation deficiencies.
- Drafts complete pull requests for missing tutorial guides and API references.
TypeScriptDiscord/Slack APIsVector DBNext.jsOctokit
"Deployed an autonomous developer advocacy agent parsing GitHub issues and Discord inquiries to generate grounded RAG solutions and draft missing technical documentation."
Action-Oriented Customer Resolution Agent
An action-oriented support agent that doesn't just converse—it actively resolves customer tickets by querying order databases, initiating refunds, and adjusting shipping parameters within authorized policy bounds.
Key Engineering Highlights:
- Integrates external database query tools for live customer order lookup.
- Strict policy guardrails enforcing monetary refund limits with automated supervisor escalation.
- Multi-turn context retention across email, chat, and CRM ticketing systems.
PythonLangChainFastAPIPostgreSQLStripe API
"Built an autonomous customer resolution agent executing transactional operations (refunds, order modifications) within strict financial bounds."
Outbound Lead Qualification & Showing Scheduler
An outbound conversational agent built in TypeScript that qualifies real estate buyer leads, answers complex property inquiries using RAG over listing brochures, and books in-person showings.
Key Engineering Highlights:
- RAG over MLS listings and architectural floorplans for accurate square footage and HOA answers.
- Two-way calendar synchronization with agent Google/Outlook calendars.
- Automated SMS reminder sequences minimizing showing no-shows.
TypeScriptNode.jsTwilio SMS/VoiceOpenAI APIPostgreSQL
"Engineered an outbound conversational real estate coordinator qualifying buyer leads via RAG over property brochures and scheduling calendar showings."
Autonomous Voice Dispatcher for Home Services
An autonomous voice dispatcher for home emergency services (plumbing, electrical, HVAC) that receives emergency calls, extracts problem severity, identifies the nearest technician via geo-routing, and dispatches appointments.
Key Engineering Highlights:
- Real-time voice transcription and emergency severity grading.
- Geo-distance routing matching technicians based on live GPS location and specialized tool inventory.
- Automated technician dispatch confirmation via SMS.
PythonFastAPITwilioPostGISGoogle Maps API
"Developed an autonomous emergency service dispatcher utilizing geo-spatial routing to match and dispatch emergency calls to field technicians."
Consumer Voice Concierge Navigating Airline & Hotel IVRs
A voice concierge agent that calls airline or hotel customer service on behalf of users, navigates complex IVR phone menus, waits on hold, and bridges the user only when a live human agent is reached.
Key Engineering Highlights:
- DTMF tone generation navigating multi-tier automated IVR phone trees.
- Hold-music and voice activity detection determining when a human representative answers.
- Seamless live three-way WebRTC bridging connecting the user.
PythonTwilio VoiceWebRTCAudio ClassificationFastAPI
"Architected a consumer voice concierge agent navigating automated IVR trees and hold queues, bridging the user only upon human agent connection."
Autonomous Statistical Analysis & Paper Extraction Agent
An autonomous research assistant that ingests messy academic papers and tabular datasets, extracts metrics, writes Python analysis scripts, executes them in isolated sandboxes, and compiles charts.
Key Engineering Highlights:
- Extracts tabular data from scientific PDFs into normalized Pandas dataframes.
- Generates and runs statistical hypothesis tests and regression models inside isolated containers.
- Automates generation of publication-ready Seaborn and Matplotlib visualization artifacts.
PythonDocker SandboxPandasSciPyMatplotlibClaude 3.5
"Built an autonomous research sandbox ingesting messy scientific publications, executing statistical hypothesis testing in Docker isolation, and generating visualization figures."
Stripe Webhook Auditor & Autonomous Dunning Workflow in Go
A multi-agent revenue recovery engine built in Go that monitors Stripe webhooks for failed subscription charges, analyzes user engagement metrics, and executes intelligent dunning workflows.
Key Engineering Highlights:
- High-throughput webhook listener processing Stripe payment failure events in Go.
- Predictive dunning schedule optimization maximizing payment recovery rates.
- Recovered over 18% of churned MRR through automated intelligent retry routing.
Go (Golang)Stripe APIPostgreSQLSendGrid APIDocker
"Engineered a revenue recovery engine in Go monitoring Stripe billing webhooks and orchestrating predictive dunning sequences, recovering 18% of failed subscription revenue."
Adversarial Consensus & Bias Elimination Framework
A multi-agent consensus system where two or more specialized agent personas debate complex dilemmas from opposing perspectives before an impartial judge agent synthesizes a balanced verdict.
Key Engineering Highlights:
- Multi-turn dialectical debate protocol eliminating single-model cognitive bias.
- Formal scoring rubrics evaluating factual consistency, logical fallacies, and empirical evidence.
- Demonstrated 22% improvement in factual accuracy on ambiguous policy evaluations.
PythonLangGraphLiteLLMPydanticFastAPI
"Designed a multi-agent debate framework deploying adversarial personas to stress-test complex strategic hypotheses, improving final synthesis accuracy by 22%."
Playwright Visual Inspector & Design System Auditor
An autonomous visual testing agent crawling web applications using Playwright, sending rendered viewports to vision LLMs to audit alignment against brand guidelines and accessibility standards.
Key Engineering Highlights:
- Automated WCAG 2.1 contrast ratio and font sizing accessibility auditing.
- Flags visual misalignment across mobile, tablet, and ultra-wide viewports.
- Generates CSS patch recommendations directly from vision model inspections.
TypeScriptPlaywrightClaude 3.7 VisionOpenCVNode.js
"Implemented an autonomous visual QA agent crawling web applications with Playwright and evaluating layout accessibility and design compliance using Claude Vision."
Hierarchical Multi-Agent Knowledge Retrieval Swarm
A hierarchical swarm of specialized retrieval agents querying disparate enterprise knowledge stores (Confluence, Notion, Google Drive, Jira) and synthesizing a unified response.
Key Engineering Highlights:
- Deconstructs broad enterprise queries into sub-queries delegated to domain-specific connectors.
- Reciprocal Rank Fusion (RRF) reranking across disparate vector indexes and keyword engines.
- Enforces source-level document permissions before response synthesis.
PythonCrewAIQdrantFastAPIPydantic
"Engineered a hierarchical knowledge retrieval swarm coordinating specialized agents across enterprise documentation silos with Reciprocal Rank Fusion reranking."
Local-First Clinical Intelligence with Offline Gemma & 2G SMS Bridge
A cross-platform, local-first clinical decision assistant designed for resource-constrained clinics. Operates offline with local Gemma models via Ollama or online with cloud Gemini models, featuring a 2G SMS bridge.
Key Engineering Highlights:
- Dynamic local-to-cloud inference switching: runs offline via Ollama (Gemma) with cloud failover to Gemini.
- Consultation pipeline: transcription -> SOAP clinical note formatting -> triage draft.
- 2G SMS fallback bridge via Africa's Talking API for messaging feature-phone patients in remote regions.
PythonFastAPIReact NativeOllama (Gemma)Google Gemini APIAfrica's Talking SMS API
"Built MedEdge, a local-first clinical assistant operating offline via Ollama (Gemma) with cloud failover to Gemini Pro for resource-constrained clinics, integrated with a 2G SMS fallback bridge for feature phones."
HIPAA-Hardened Patient Record Vault with pgvector & Immutable Logs
A high-security, HIPAA-compliant patient record repository featuring document QA search using vector embeddings, role-based access control (RBAC), and database-level immutable audit logging.
Key Engineering Highlights:
- Vector search RAG using pgvector inside PostgreSQL with HNSW index optimization.
- Database-level immutable audit logging enforced via PostgreSQL triggers to prevent tampering.
- Temporary AWS S3 pre-signed URLs guaranteeing zero direct-link document exposure.
Next.jsNode.jsTypeScriptPostgreSQL (pgvector)AWS S3 Pre-Signed URLsTailwind CSSDocker
"Architected a HIPAA-compliant healthcare document vault featuring pgvector semantic Q&A, immutable PostgreSQL audit triggers, and temporary AWS S3 signed URLs guaranteeing zero unauthorized data exposure."
Natural Language Shift Scheduling with OpenAI Tool Use
An AI-powered care scheduling platform that enables agency coordinators to manage shifts and enforce labor compliance using natural language prompts, offloading validation to an agentic tool loop.
Key Engineering Highlights:
- OpenAI function calling translating natural language staffing directives into optimal shift assignments.
- Real-time certification verification and automatic double-booking prevention enforced via PostgreSQL constraints.
- Containerized multi-service deployment with Next.js UI, Express API, and PostgreSQL.
Next.jsNode.jsTypeScriptExpressPostgreSQLOpenAI Function CallingDocker
"Developed an AI shift scheduling platform for care agencies translating natural language requests into compliant staffing rosters via OpenAI tool use, preventing double-bookings and certification lapses."
Guided Socratic Dialogue Platform (Google Gemma Hackathon)
An AI-driven educational platform designed for the Google Gemma 4 Good Hackathon, bridging the gap between personalized tutoring and resource-constrained environments by running local Gemma models on edge devices.
Key Engineering Highlights:
- On-device guided Socratic dialogue using Ollama and Gemma.
- Hybrid edge-to-cloud architecture supporting offline rural school classrooms.
- Admin telemetry panel tracking student mastery curves.
React NativeFastAPIOllamaGoogle GemmaWebSockets
"Engineered Socrates, an educational platform conducting guided Socratic dialogues on edge devices using local Gemma models, built for the Google Gemma Hackathon to support resource-constrained classrooms."
Next.js 16, FastAPI & Gemini 2.5 Pro Vision 7-Agent Network
An AI-powered film studio operating system and multi-agent network automating video ingestion, screenplay analysis, casting performance, and auto-dubbing translation.
Key Engineering Highlights:
- 7 specialized agents coordinating video ingestion, pre-production breakdown, and continuous takes analysis.
- Audition analysis evaluating pitch, tempo, and emotional intensity using Gemini 2.5 Pro Vision.
- Unified backend API Gateway (FastAPI) serving a Next.js web panel and an Expo mobile app.
Next.js 16ReactFastAPIExpo / React NativeGemini 2.5 Pro VisionPostgreSQLDocker
"Architected Studio OS, an AI film production platform coordinating 7 specialized agents to automate screenplay breakdown and audition analysis using Gemini 2.5 Pro Vision."
Real-Time Multiplayer Game & Media Synchronization Platform
A production real-time multiplayer card gaming platform powered by Node.js, Express, WebSockets, React, Zustand, and Neon PostgreSQL, packaged for native iOS and Android via Capacitor.
Key Engineering Highlights:
- Authoritative game server handling concurrent trick resolution and turn timers over WebSockets (<50ms latency).
- State management with Zustand and optimistic UI rendering.
- Cross-platform distribution across Web, iOS, and Android via Capacitor.
Node.jsExpressWebSocketsReactZustandDrizzle ORMNeon PostgreSQLCapacitor
"Engineered a real-time multiplayer card gaming platform supporting sub-50ms game state synchronization over WebSockets, packaged for native iOS and Android via Capacitor."
Dockerized Full-Stack Card Gaming Engine
A containerized web application platform delivering multiplayer card game rule engines, player matchmaking, and responsive frontend canvas components.
Key Engineering Highlights:
- Complete rule validation for trick-taking multiplayer card games.
- Reverse-proxied with Nginx and containerized with Docker.
- Low-latency state broadcasting to active browser sessions.
Node.jsReactDockerNginxPostgreSQL
"Delivered a Dockerized card game platform with Nginx reverse proxying and WebSocket state management."
Deterministic Game Physics & Board Logic Engine in Python
A classic multiplayer Ludo board game implementing complete rules, piece collision logic, safe squares, and AI bot opponents.
Key Engineering Highlights:
- Deterministic collision detection and board coordinate mappings.
- Heuristic AI bot opponents simulating strategic token movements.
- Built with Pygame showcasing low-level event loops and graphics rendering.
PythonPygame
"Implemented a classic multiplayer board game engine in Python with deterministic piece collision rules and bot logic."
React Native, Expo, NativeWind, Prisma & Supabase
A modern cross-platform mobile quiz application built using React Native, Expo, TypeScript, TailwindCSS (NativeWind), Prisma, and Supabase.
Key Engineering Highlights:
- NativeWind mobile styling with fluid animations and responsive mobile layouts.
- Offline question caching and background synchronization with Supabase.
- Comprehensive mobile test coverage using Jest and React Native Testing Library.
React NativeExpo RouterTypeScriptNativeWindPrismaSupabaseJest
"Shipped a cross-platform mobile quiz app using React Native, Expo, and NativeWind with Supabase synchronization, upholding strict mobile UI test coverage."
Digital Footprint Growth Platform & Personal CRM
A private, single-user digital footprint growth platform acting as an intelligent personal CRM, PR outreach tracker, conference networking workspace, and product launch planner.
Key Engineering Highlights:
- Next.js App Router with Server Components and PostgreSQL database adapter (@prisma/adapter-pg).
- Automated touchpoint tracking and interaction graph visualization.
- Tailwind CSS responsive design with dark mode support.
Next.js (App Router)ReactTypeScriptPrisma ORMPostgreSQLTailwind CSS
"Engineered FootprintOS (Next.js, Prisma, PostgreSQL), a personal CRM and PR automation tool tracking technical outreach, conference touchpoints, and product launches."
Distributed Microservices, Serverless Pipelines & Enterprise RAG
A complete enterprise cloud suite demonstrating distributed microservices, event-driven serverless architectures, and generative AI copilot systems built with C# (.NET 8), Apache Kafka, AKS, and Azure Cloud.
Key Engineering Highlights:
- Cloud-native logistics microservices with Kafka event streaming, CQRS, MediatR, and YARP API Gateway.
- Event-driven serverless document processing using Azure Functions (.NET 8 Isolated Worker), Cosmos DB, and SignalR.
- Enterprise RAG system built with Microsoft Semantic Kernel and Azure OpenAI.
C# (.NET 8)Apache KafkaAzure Kubernetes Service (AKS)YARPAzure FunctionsCosmos DBSemantic KernelBicep
"Delivered the Azure Enterprise Cloud Suite featuring .NET 8 event-driven microservices on Kafka/AKS, serverless Azure Functions pipelines, and enterprise RAG with Microsoft Semantic Kernel."
Agentic Lakehouse, Streaming Anomaly & DataOps RAG
Enterprise-grade AI & Data Architecture showcase projects demonstrating production Agentic Lakehouses, Real-time Streaming Anomaly Detection, and DataOps RAG pipelines.
Key Engineering Highlights:
- Agentic Lakehouse combining Medallion Delta Lake architecture with autonomous data quality agents.
- Real-time streaming anomaly detection engine built with Apache Kafka and Flink.
- DataOps RAG pipeline establishing automated data ingestion, validation, and evaluation tracking.
Apache SparkApache KafkaApache FlinkDelta LakePythonFastAPIDocker
"Authored enterprise architecture blueprints for Agentic Lakehouses and streaming anomaly detection engines handling high-volume real-time events across distributed data pipelines."
ML Competition Pipeline & Deep Stacking Workbench
A machine learning competition engineering workbench with automated feature engineering, Bayesian hyperparameter optimization, and multi-model ensembling.
Key Engineering Highlights:
- Automated stacking combining LightGBM, XGBoost, and PyTorch deep neural networks.
- Optuna Bayesian hyperparameter optimization tracking cross-validation scores.
- Feature importance and error analysis dashboards.
PythonPyTorchLightGBMXGBoostOptunaScikit-LearnPandas
"Built an automated ML competition pipeline featuring Bayesian hyperparameter optimization (Optuna) and multi-model neural network stacking across tabular and multimodal datasets."
16-Notebook Systems Engineering & Cloud Architecture Guide
A comprehensive 16-notebook technical suite covering advanced Python paradigms, concurrency models, distributed algorithms, memory management, and cross-cloud architectural mappings.
Key Engineering Highlights:
- Exhaustive coverage of OOD principles, concurrency patterns, and distributed algorithms.
- Comparative technical mappings of architectural primitives across AWS, GCP, and Azure.
- Memory optimization and low-level performance-tuning strategies for core Python architectures.
PythonJupyter NotebooksCloud ArchitectureDistributed Algorithms
"Authored a 16-notebook engineering suite covering system design patterns, distributed concurrency models, and comparative AWS/GCP/Azure cloud architecture mappings."
Clean Architecture & Design Smell Refactoring Laboratories
Code laboratories demonstrating GoF design patterns (Creational, Structural, Behavioral), clean architecture principles, and refactorings of common architectural smells.
Key Engineering Highlights:
- Comprehensive student packs and labs covering SOLID principles and system design.
- Refactoring case studies transforming legacy antipatterns into clean decoupled architectures.
- Complete UML diagrams and reference implementations across Python and TypeScript.
PythonTypeScriptUMLClean ArchitectureDesign Patterns
"Developed clean code laboratories demonstrating GoF design patterns, SOLID principles, and architectural refactoring strategies."
Autonomous SEO & Content Generation Agent
An autonomous AI-powered content creation system designed to automate technical blog post generation, SEO keyword research, knowledge retrieval, and audio asset synthesis.
Key Engineering Highlights:
- Autonomous SEO keyword clustering and search intent analysis.
- Grounded technical drafting utilizing codebase RAG context.
- Automated text-to-speech audio synthesis for accessible audio articles.
PythonFastAPIOpenAI APITTS Audio ModelsMarkdown Ingestion
"Created BlogAgent, an autonomous content generation agent performing SEO research, RAG-grounded drafting, and automated audio synthesis."
Automated Video Search, Transcript Ingestion & Topic Intelligence
An automated video intelligence agent performing daily targeted YouTube searches, extracting closed captions, and generating executive summaries of emerging AI research topics.
Key Engineering Highlights:
- Scheduled daily search and video discovery across designated AI research topics.
- Automated transcript extraction and token-efficient summarization.
- Generates daily digest newsletters and trend alerts.
PythonYouTube Data APIYouTube Transcript APIDockerClaude API
"Built an automated video intelligence agent extracting and analyzing YouTube transcripts to synthesize executive trend reports on emerging AI research."
Enterprise Image & Video Diffusion Pipeline with Zero-Allocation Transcoding
An enterprise GenAI image & video generation studio suite engineered with Go 1.26, Next.js 16, Expo SDK 56, and Redis Streams, delivering zero-allocation ring buffers, warm-cache GPU affinity routing, dynamic admission backpressure, and zero-disk streaming AV1 transcoding.
Key Engineering Highlights:
- Turborepo monorepo uniting Next.js 16 web studio, Expo SDK 56 mobile app, Go 1.26 media pipeline, and shared GraphQL codegen packages.
- Zero-disk streaming transcoding engine via FFmpeg 7.x AV1 with dynamic admission backpressure and queue priority arbitration.
- Warm-cache GPU worker affinity routing reducing cold-start diffusion model load overhead by 80%.
- Real-time GraphQL subscriptions and AsyncAPI event bus over Redis 7.4 Streams for distributed generation state machine sync.
TurborepoGo 1.26Next.js 16Expo SDK 56Redis StreamsFFmpeg AV1GraphQLThree.jsDocker
"Architected StudioForge, an enterprise GenAI diffusion studio Turborepo (Go 1.26, Next.js 16, Expo SDK 56, Redis 7.4 Streams) featuring zero-allocation ring buffers, GPU affinity routing, and real-time streaming AV1 transcoding."
Real-Time Incident Commander & GPU Fleet Triage for High-Throughput GenAI Clusters
An autonomous GPU inference cluster observability and incident triage platform engineered for ultra-scale GenAI infrastructure ($35M+ ARR scale), featuring sub-millisecond telemetry ingestion, zero-allocation memory pools, and live incident war rooms.
Key Engineering Highlights:
- High-throughput Go 1.25 metric stream ingester with zero-allocation buffers (sync.Pool) handling 100K+ metrics/sec.
- PostgreSQL 17 TimescaleDB hypertable partitioning for sub-second analytical queries across millions of GPU telemetry points.
- Real-time incident war room with live multi-agent triage logs streaming via GraphQL Subscriptions to Web and Mobile.
- Autonomous root cause isolation correlating NVLink degradation, thermal throttling, and VRAM memory fragmentation.
Go 1.25TimescaleDBPostgreSQL 17Next.js 16GraphQLOpenTelemetryDocker
"Architected PulseOps AI, an autonomous GPU inference fleet triage platform in Go 1.25 and TimescaleDB with zero-allocation buffers, handling 100K+ metrics/sec and streaming live incident mitigation trees via GraphQL Subscriptions."
Sub-Second Clinical Audio Slicing & FHIR-Compliant SOAP Encounter Generation
A high-throughput clinical ambient scribe and multimodal processing platform executing sub-millisecond silence boundary slicing, parallel GPU Whisper v3 Turbo fan-out, and FHIR-compliant clinical SOAP note generation.
Key Engineering Highlights:
- Low-latency audio chunking daemon in Go with FFmpeg silence boundary detection slicing encounters in under 5ms.
- 60x parallel Whisper v3 Turbo fan-out architecture transcribing a 30-minute encounter in under 4 seconds.
- Smith-Waterman overlap boundary alignment eliminating hallucinated boundary artifacts between chunk boundaries.
- HIPAA-compliant, structured FHIR JSON output and reactive SOAP note review workspace with clinician audit trails.
Go 1.25Whisper v3 TurboNext.js 16Expo SDK 56FFmpegFHIRHIPAADocker
"Engineered MediScribe AI, a clinical audio processing platform achieving 60x parallel Whisper v3 Turbo fan-out (<4s for 30m encounter) with Smith-Waterman overlap alignment and HIPAA-compliant FHIR SOAP generation."
Enterprise Distributed Web Scraping, Proxy Arbitration & Ingestion Pipeline
An enterprise-grade distributed web scraping engine comprising a NestJS API gateway, high-concurrency Go worker daemons, and Redis queue topology with Kubernetes HPA auto-scaling, handling high-volume page crawling with dynamic rate limiting.
Key Engineering Highlights:
- Modular microservices architecture separating NestJS API ingest, Valkey/Redis task queue, and Go worker scrapers.
- High-concurrency Go worker runtime converting complex HTML DOM trees to clean Markdown in sub-10ms.
- Dynamic per-domain token bucket rate limiting, automated proxy rotation, and exponential backoff retry policies.
- Kubernetes Horizontal Pod Autoscaler (HPA) scaling worker fleet based on queue backlog depth and latency SLAs.
GoNestJSRedisKubernetesDockerTypeScriptMarkdown
"Architected CrawlPulse, an enterprise distributed web scraping engine featuring NestJS, high-concurrency Go workers, and Redis queue topology with Kubernetes HPA autoscaling and sub-10ms Markdown DOM conversion."
Enterprise Crisis Response, Audit Trail & Incident Case Management Microservices
A high-reliability emergency response and case management distributed platform engineered with Spring Boot, Keycloak OIDC/OAuth2 security, audit service streaming, and immutable incident event ledgers.
Key Engineering Highlights:
- Domain-driven microservices architecture decoupling case management, audit event logging, and emergency dispatch.
- Centralized Keycloak realm authentication with granular RBAC for first responders, dispatchers, and medical coordinators.
- High-throughput asynchronous audit logging daemon capturing cryptographic SHA-256 state transitions for regulatory compliance.
- Containerized multi-service deployment with Docker Compose, automated health probes, and PostgreSQL connection pooling.
Java 21Spring BootKeycloakPostgreSQLDockerMicroservicesMaven
"Architected CrisisHub, an emergency dispatch and crisis case management microservices platform with Spring Boot, Keycloak OIDC RBAC, and immutable audit service streaming for mission-critical disaster response."
Sub-Second IoT Telemetry Gateway & Geofence Alerting Microservices
A real-time IoT device telemetry and geospatial asset tracking platform featuring sub-second GPS coordinate ingestion, polygon geofence evaluation, and automated alert dispatch.
Key Engineering Highlights:
- High-throughput ingestion service processing streaming GPS coordinates and sensor payloads from distributed hardware devices.
- Sub-millisecond point-in-polygon geofencing engine triggering real-time enter/exit alerts across dynamic geographical boundaries.
- Keycloak OAuth2 device authorization protecting device telemetry endpoints and management APIs.
- Time-series geospatial queries optimized for trajectory replay and asset utilization analytics.
GoJavaKeycloakTimescaleDBPostgreSQLDockerIoTGeospatial
"Engineered GeoPulse, a real-time IoT fleet telemetry platform featuring high-concurrency coordinate ingestion, sub-millisecond point-in-polygon geofence evaluation, and Keycloak device authorization."
Full-Duplex Grounded Video/Audio Voice Agent with Timestamp Citation Seeking
An interactive voice intelligence platform built for the AssemblyAI Voice Agent Hackathon that turns long-form videos, webinars, and podcasts into full-duplex conversational partners with spoken answers grounded in exact transcript timestamps.
Key Engineering Highlights:
- Combines AssemblyAI asynchronous STT, speaker diarization, auto-chapters, and LeMUR (Claude reasoning) with real-time Voice Agent WebSockets.
- Browser AudioWorklet circular ring buffers with instant playback halt flags on speech onset for zero-echo conversational barge-in.
- Deterministic search_transcript tool calling that grounds every spoken response in verified transcript segments with click-to-seek video synchronization.
- Sub-300ms turnaround latency with live latency HUD telemetry and a 20-assertion offline test suite.
AssemblyAI Voice AgentLeMUR ClaudeAudioWorkletWebSocketsTypeScriptNode.jsSpeaker Diarization
"Built VoxDive, a full-duplex voice agent over video and podcast archives using AssemblyAI STT, speaker diarization, LeMUR Claude tool calling, and Web Audio API AudioWorklet buffers for sub-300ms grounded spoken answers with timestamp citations."
Outbound AWS Chime SIP Incident Paging, Spoken Command Triage & Multi-Agent ECS Rollback Swarm
An autonomous P1 outage response system built for the AWS Cloud Development Series Agentic AI Hackathon that replaces silent push alerts with interactive outbound AWS Chime SIP phone calls and multi-agent remediation via the AWS Strands A2A protocol.
Key Engineering Highlights:
- Places outbound phone calls via AWS Chime Voice Connector SIP trunking to brief on-call engineers and parse spoken natural-language remediation commands.
- Coordinates a 4-agent swarm over the open AWS Strands Agent-to-Agent (A2A) protocol: TriageSupervisorAgent, TelephonyVoiceAgent, DevOpsRemediationAgent, and IncidentReporterAgent.
- Executes automated canary traffic drains and Amazon ECS service rollbacks in seconds upon voice confirmation.
- Synthesizes structured executive post-mortem Markdown logs with 100% unit test verification (7/7 test suite passing).
AWS Strands A2AAWS Chime SIPAmazon BedrockClaude 3.5 SonnetTypeScriptAmazon ECSCloudWatch
"Architected DispatchAgent.AI, an autonomous incident commander integrating AWS Chime SIP voice trunking, Amazon Bedrock tool use, and the AWS Strands A2A multi-agent protocol to execute spoken-command ECS rollbacks and automated post-mortems."
Zero-Dependency Session Search, Mermaid Edit-Sequence Generator & Subagent Swarm Tracker
A lightweight, zero-dependency observability dashboard and file modification visualizer for Google Antigravity AI agent sessions that parses multi-thousand-line JSONL trajectories into interactive timelines and Mermaid flowcharts.
Key Engineering Highlights:
- Built 100% on Python's standard library (http.server, sqlite3, json, pathlib) with zero external pip dependencies and instant desktop app mode launcher.
- Indexes conversation_summaries.db and transcript.jsonl to render chronological file edit sequences ([NEW] -> [EDIT] -> [READ]) and automated Mermaid flowcharts.
- In-app syntax-highlighted code reviewer with 1-click 'Open in VS Code' (exact line jump) and 'Reveal in File Explorer' OS hooks.
- Subagent Swarm tracker inspecting invoke_subagent delegations, tool error distributions, universal drag-and-drop JSONL parsing, and secret-scrubbed trace export.
Python 3.12+SQLiteMermaid.jsHTML5 / CSS3Zero-DependencyAgent ObservabilityVS Code Hooks
"Developed Antigravity Transcript Visualizer, a zero-dependency Python observability dashboard indexing SQLite agent summaries and JSONL traces into Mermaid file-edit sequence diagrams, subagent swarm trees, and 1-click VS Code line jumpers."
Local-First CrewAI Scraper with JA4 TLS Impersonation, Selectolax C-Parser & Stealth Playwright
A distributed, local-first multi-agent web crawling and structured data extraction platform engineered to bypass modern anti-bot WAFs (Cloudflare, Akamai, DataDome), hydrate dynamic SPAs, and enforce self-healing Pydantic v2 schemas.
Key Engineering Highlights:
- 3-agent CrewAI pipeline: Reconnaissance Scout Agent (WAF/header probing), Stealth Browser Navigator Agent (interactive SPA hydration), and Data QA Validator Agent.
- Low-level curl_cffi HTTP client spoofing Chrome JA3/JA4 TLS handshakes and HTTP/2 frame parameters alongside a selectolax C-engine HTML parser (30x faster than BS4).
- Pluggable LLM router toggling seamlessly between $0 local GPU Ollama models (qwen2.5-coder, llama3.1) and Google Gemini 2.5 Flash.
- SQLite WAL persistence with an automated Dead-Letter Queue (DLQ) capturing DOM dumps and stack traces for self-healing schema recovery.
Python 3.13CrewAIPlaywright Stealthcurl_cffi (JA4 TLS)selectolaxPydantic v2SQLite WALOllama / Gemini
"Architected AgentCrawl, an autonomous CrewAI web extraction platform combining curl_cffi JA4 TLS fingerprint spoofing, selectolax C-engine DOM parsing, stealth Playwright SPA navigation, and SQLite WAL Dead-Letter Queues."
High-Throughput Model Context Protocol (MCP) Swarm & FP8 vLLM-ROCm Inference Engine
A high-performance Model Context Protocol (MCP) multi-agent orchestration engine built for the AMD Developer Hackathon (ACT III), optimized for AMD Instinct MI300X GPU accelerators and ROCm 6.2 high-throughput inference.
Key Engineering Highlights:
- Dynamic MCP tool registry allowing autonomous subagent swarms to discover and invoke ROCm tensor auditors and AST grounding validators.
- Optimized for AMD Instinct MI300X (ROCm 6.2) with FP8 quantized vLLM inference delivering 284.5 tokens/sec and 38.2ms Time-to-First-Token (TTFT).
- Deterministic execution trace verifier asserting 100% AST and tool-call provenance across multi-agent delegations.
- Strict TypeScript implementation with interactive telemetry dashboard and 100% passing offline test suite.
TypeScriptAMD ROCm 6.2Instinct MI300XModel Context ProtocolvLLM FP8Multi-Agent Mesh
"Engineered OmniMesh-ROCm, an enterprise Model Context Protocol (MCP) multi-agent orchestration mesh optimized for AMD Instinct MI300X and ROCm 6.2 FP8 inference (284.5 tok/s, 38.2ms TTFT) with deterministic AST trace grounding."
Real-Time Cluster Anomaly Detection, Autonomous Instance Scaling & Serverless Worker Dispatch
An autonomous multi-agent CloudOps infrastructure orchestrator built for the Vultr Agent Rush Hackathon that monitors compute clusters, detects CPU/memory anomalies, and dispatches serverless workers and instance resizing without human intervention.
Key Engineering Highlights:
- Multi-agent coordination loop pairing a Telemetry Anomaly Detection Subagent with an Autonomous CloudOps Remediation Engine.
- Automated Vultr compute actions including vertical instance resizing (vc2-8c-16gb) and horizontal serverless worker container spawning.
- Immutable CloudOps audit log recording anomaly triggers, confidence scores, and remediation execution latencies.
- Zero-dependency TypeScript architecture with real-time cluster health HUD and deterministic test suite.
TypeScriptVultr Cloud APIMulti-Agent SwarmCloudOpsAuto-ScalingSRE Automation
"Built VultrSwarm-Agentic, an autonomous multi-agent CloudOps orchestrator that detects compute telemetry anomalies and executes automated Vultr instance scaling and serverless worker dispatch with immutable audit logging."
Bedrock Converse API IaC Auditor, Graviton3 FinOps Modeler & Zero-Downtime Git Patch Synthesizer
An autonomous cloud governance and FinOps engine built for the Amazon Developer Hackathon using the Amazon Bedrock Converse API to audit Terraform/OpenTofu HCL manifests against the AWS Well-Architected Framework and synthesize reviewable Git remediation diffs.
Key Engineering Highlights:
- Built on the AWS Bedrock Converse API standard with Claude 3.5 Sonnet v2 cross-region inference profiles and strict JSON Schema tool calling (audit_iac_manifest, estimate_finops_savings, synthesize_remediation_diff).
- AST-tolerant HCL lexer detecting unrestricted 0.0.0.0/0 SSH/RDP ingress, unencrypted storage volumes, and oversized x86 instances without requiring a live Terraform binary.
- Projects annualized FinOps savings ($4,080/yr per cluster benchmarked) and AWS Graviton3 carbon offset reductions.
- Synthesizes zero-downtime unified Git patches with a side-by-side diff viewer and companion sample IaC repository (cloudpulse-sample-iac).
Amazon BedrockClaude 3.5 SonnetConverse APITypeScriptTerraform HCLOpenTofuFinOps
"Architected CloudPulse AI using the Amazon Bedrock Converse API and Claude 3.5 Sonnet to audit Terraform/OpenTofu manifests across Security, FinOps, and Sustainability pillars, identifying $4,080/yr cluster savings and generating unified Git remediation patches."
Deterministic Git Diff & AST Grounded Release Note & Pull Request Auditor
An evidence-first multi-agent system built for the micro1 AI Hackathon that audits software release notes and pull request reviews by grounding every single finding in physical Git commits, diffs, and AST changes before alerting maintainers.
Key Engineering Highlights:
- Router Agent classifies incoming repository queue items and delegates to specialized auditors (Specialist G for Changelogs, Specialist E for PR Reviews).
- Eliminates LLM hallucinations by requiring deterministic Git commit hash and line-level diff verification prior to surfacing findings.
- Achieved 1.00 Precision (zero false alarms) vs 0.70 baseline and 0.87 Mean F1 score across live benchmark runs.
- Multi-provider LLM routing (OpenAI, Anthropic, Groq, OpenRouter) with mandatory human-in-the-loop approval gates.
TypeScriptPythonMulti-Agent SwarmGit Diff GroundingAST VerificationOpenAI / ClaudeCI Quality Gate
"Engineered Triage Inbox, an evidence-first multi-agent GitHub PR and release auditor that grounds every claim in physical Git diffs and AST changes, achieving 1.00 precision (zero false alarms) and 0.87 F1 vs 0.70 baseline."
Cross-Silo Enterprise Knowledge Graph with Weighted Reciprocal Rank Fusion (WRRF k=60)
An enterprise Hybrid GraphRAG knowledge mesh built for the Galuxium Nexus V2 Hackathon that unifies fragmented SaaS silos (Jira tickets, GitHub PRs, Datadog APM traces, and Workday org graphs) to perform multi-hop causal root-cause investigation.
Key Engineering Highlights:
- Typed enterprise knowledge graph connecting Incidents -> Affected Services -> Merged PRs -> Commit Authors across corporate data silos.
- Dual-retrieval pipeline fusing dense cosine vector search with degree-constrained multi-hop breadth-first graph traversal via Weighted Reciprocal Rank Fusion (WRRF k=60).
- Demonstrates 3-hop causal root-cause discovery from production outage alarm to offending pull request in under 45ms with 100% provenance verification.
- TypeScript implementation with 5/5 passing unit assertions and zero hallucinated citations.
TypeScriptHybrid GraphRAGWRRF Fusion (k=60)Multi-Hop BFSVector SearchKnowledge GraphRoot-Cause Analysis
"Built NexusGraph Enterprise, a Hybrid GraphRAG engine uniting Jira, GitHub, and Datadog silos via Weighted Reciprocal Rank Fusion (WRRF k=60) of dense vector search and 3-hop graph traversal for sub-45ms root-cause lineage."
Sub-50ms TTFT Visual Grounding & Zero-Drift OS Desktop Automation via Nebius & NVIDIA NIM
An autonomous OS-level Robotic Process Automation (RPA) platform powered by Nebius Token Factory and NVIDIA NIM on H100 SXM GPU clusters, grounding natural-language instructions into 1920x1080 screen coordinates for zero-drift GUI automation without APIs.
Key Engineering Highlights:
- Dual-engine inference orchestrator running Meta-Llama-3.1-70B-Instruct on Nebius H100 80GB SXM with automatic failover to NVIDIA NIM (Llama-3.3-70B).
- FP8 TensorRT-LLM quantization delivering sub-50ms Time-to-First-Token (TTFT) and 186.5 tokens/sec sustained throughput (>70% latency reduction vs FP16).
- Visual Grounding Engine computing UI element centroids with strict spatial safety bounding boxes to prevent out-of-bounds OS clicks.
- 100% passing test suite verifying coordinate grounding, failover routing, and multi-step desktop action sequences.
TypeScriptNebius Token FactoryNVIDIA NIMTensorRT-LLM FP8H100 GPULlama 3.1 70BVisual Grounding
"Engineered TensorMesh RPA, an OS-level visual desktop automation agent powered by Nebius Token Factory and NVIDIA NIM FP8 TensorRT-LLM on H100 GPUs, achieving sub-50ms TTFT and 186.5 tok/s with spatial boundary safety assertions."
30+ FPS OpenCV ONNX Spatial Hazard Detection, HSV PPE Verification & AWS S3/SNS Alerting
An industrial edge computer vision safety monitoring pipeline built for the OpenCV AI Competition 2026 (powered by AWS) that analyzes RTSP camera feeds at 30+ FPS to detect restricted danger-zone intrusions and PPE compliance violations.
Key Engineering Highlights:
- Real-time edge detection using OpenCV cv2.dnn ONNX object detectors paired with HOG pedestrian fallback and pinhole camera depth estimation.
- Polygonal hazard zone assertion via cv2.pointPolygonTest and high-visibility PPE vest inspection in HSV color space (cv2.inRange).
- Temporal persistence debounce filter requiring multi-frame confirmation to eliminate false alarms from camera flicker or brief occlusions.
- Automated incident snapshot dispatch to AWS S3 and real-time push alerting via Amazon SNS with full offline synthetic-frame test coverage.
Python 3.10+OpenCV cv2.dnnONNX / YOLOAWS S3 / SNSSpatial GeometryHSV PPE InspectionEdge AI
"Developed EdgeSentry, a 30+ FPS edge computer vision workplace safety monitor using OpenCV cv2.dnn ONNX detection, pointPolygonTest spatial hazard checks, HSV PPE verification, temporal debounce filtering, and AWS S3/SNS alerting."
Audio Denoising, Few-Shot Voice Cloning, Talking Portrait Animation & Guardian PIN Gate
A child-safe generative AI mobile and backend platform enabling bereaved families to transform preserved WhatsApp voice notes and portrait photos into guardian-approved bedtime stories and milestone tribute videos.
Key Engineering Highlights:
- 4-stage generative media pipeline: DeepFilterNet3 + Silero VAD audio restoration -> F5-TTS / ElevenLabs few-shot voice cloning -> LivePortrait / Hedra talking portrait animation -> FFmpeg 7 watermarking.
- Mandatory Guardian Review Gate requiring parental biometric or 4-digit PIN approval before any synthesized video enters the child's storybook player.
- Audio SNR and harmonic clarity analyzer scoring noisy mobile recordings (.m4a, .ogg) prior to voice synthesis.
- React Native (Expo SDK 52) mobile app paired with an asynchronous FastAPI / Pydantic v2 backend and encrypted S3/R2 media vault.
React NativeExpo SDK 52FastAPIPython 3.12+DeepFilterNet3F5-TTS / ElevenLabsLivePortraitFFmpeg 7PostgreSQL
"Architected Memorial Storyteller, a child-safe generative audio/video platform combining DeepFilterNet3 speech enhancement, few-shot voice cloning, LivePortrait facial animation, and a PIN-protected Guardian Review Gate on Expo SDK 52 and FastAPI."
Zero-GPU Symbolic Program Synthesis & Exact-Match Grid Reasoning (91.7% Benchmark Solve Rate)
A neuro-symbolic program-synthesis engine built for the Kaggle ARC Prize 2026 Code Competition that solves abstract visual reasoning tasks by searching compositions of pure NumPy Domain-Specific Language (DSL) grid transformations.
Key Engineering Highlights:
- Rich NumPy DSL primitive library covering dihedral rotations, reflections, transpositions, gravity projection, flood-fill object extraction, periodic tiling, and bijective color re-mapping.
- Depth-first iterative deepening program synthesizer with intermediate grid state hashing and pruning to eliminate redundant sub-trees.
- Exact-match verification gate requiring 100% pixel identity across all demonstration pairs before emitting candidate test predictions.
- Achieved a 91.7% solve rate (88/96 benchmark tasks) running purely on CPU with zero GPU dependency and valid Kaggle submission.json export.
Python 3.10+NumPy DSLProgram SynthesisARC-AGISymbolic SearchKaggle Competition
"Engineered a Neuro-Symbolic ARC-AGI Solver using pure NumPy DSL program synthesis and state-hashed depth-first search, achieving a 91.7% solve rate (88/96 benchmark tasks) with exact demonstration verification and zero GPU overhead."
Anisotropic 3D Kalman Filtering, Hungarian Optimal Data Association & Mitosis Lineage Tracking
A 3D developmental microscopy cell segmentation and spatiotemporal lineage tracking pipeline built for the CZ Biohub Kaggle Challenge, tracking dividing cells across volumetric time-lapses with zero ID switches.
Key Engineering Highlights:
- 3D connected-component volumetric segmentation extracting anisotropic (z, y, x) cell centroids, bounding volumes, and intensity profiles.
- Constant-velocity 6-state 3D Kalman Filter predicting inter-frame cellular motion trajectories across microscopy time steps.
- Optimal global frame-to-frame assignment via the Hungarian algorithm (scipy.optimize.linear_sum_assignment) with gated Euclidean/anisotropic cost matrices.
- Automated mitosis (cell division) parent-daughter lineage linker exporting standardized Cell Tracking Challenge (CTC res_track.txt) archives with 1.0 purity.
PythonSciPy Hungarian3D Kalman FilterVolumetric MicroscopyCTC Lineage FormatBioinformatics
"Built a 3D Spatiotemporal Cell Tracker for developmental microscopy using 6-state anisotropic Kalman filtering, Hungarian optimal data association, and automated mitosis lineage reconstruction, achieving 1.0 tracking purity and 0 ID switches."
Multi-Plane 2.5D ResNet-18 Volumetric MRI Fusion & Quadratic Weighted Kappa Calibration
A multi-sequence volumetric knee MRI classification pipeline built for the RSNA Kaggle Competition to grade ACL tears, meniscal injuries, and cartilage defects across Sagittal, Coronal, and Axial planes.
Key Engineering Highlights:
- Volumetric DICOM/NumPy preprocessing pipeline performing centre-of-mass slice selection, spatial resampling, and percentile z-score intensity normalization.
- 2.5D multi-view ResNet-18 architecture fusing adjacent volumetric slices across Sagittal T2, Coronal PD, and Axial fat-suppressed MRI sequences.
- End-to-end threshold calibration optimizing directly for the competition's Quadratic Weighted Kappa (QWK) ordinal grading metric.
- Deterministic inference pipeline outputting calibrated multi-class severity probabilities to Kaggle submission.csv.
PythonPyTorch2.5D ResNet-18Volumetric MRIDICOM ProcessingQuadratic Weighted Kappa
"Developed a Multi-Sequence 2.5D Volumetric MRI Classifier in PyTorch for the RSNA Knee Abnormality Detection challenge, fusing Sagittal, Coronal, and Axial slices with Quadratic Weighted Kappa (QWK) threshold calibration."
Stochastic Forward-Model Agricultural Digital Twin & UCT Tree Search Policy Engine
An autonomous agricultural planning agent built for the Kaggriculture Kaggle Simulation Competition that maximizes seasonal crop yield, revenue, and soil sustainability under stochastic weather and market dynamics.
Key Engineering Highlights:
- Cloneable digital-twin forward simulation environment modeling soil moisture depletion, NPK nutrient balance, rainfall stochasticity, and crop growth stages.
- Upper Confidence Bound for Trees (UCT) Monte-Carlo Tree Search (MCTS) planner evaluating multi-step irrigation, fertilization, and harvest policies.
- Time-capped 40-iteration rollout horizon with domain-informed heuristic pruning to meet strict competition per-turn compute budgets.
- Multi-objective reward shaping balancing net seasonal revenue against long-term soil health and water conservation penalties.
PythonMCTS UCTDigital TwinStochastic SimulationSequential Decision AIKaggle
"Engineered Kaggriculture MCTS, an autonomous agricultural planning agent pairing a cloneable digital-twin crop simulator with Upper Confidence Bound for Trees (UCT) Monte-Carlo Tree Search to optimize yield under stochastic weather."
Autonomous Vendor Evaluation, Contract RAG & SQL Spend Analytics Dashboard
A full-stack multi-agent enterprise procurement workspace combining a FastAPI backend, interactive Streamlit executive portal, LangChain agent orchestration, SQLite transactional spend analytics, and ChromaDB contract RAG.
Key Engineering Highlights:
- Multi-agent procurement workflow coordinating vendor risk scoring, RFP clause comparison, and purchase order compliance checks.
- Semantic vector retrieval over vendor master service agreements (MSAs) and SLA documents using ChromaDB embeddings.
- Natural-language SQL spend analytics querying historical vendor pricing, delivery SLAs, and category budgets stored in SQLite.
- Interactive Streamlit operations dashboard with real-time agent reasoning traces and downloadable audit reports.
PythonFastAPIStreamlitLangChainChromaDB RAGSQLiteProcurement AI
"Built an Enterprise Procurement Agent workspace integrating FastAPI, LangChain, ChromaDB contract RAG, SQLite spend analytics, and a Streamlit operations portal to automate vendor evaluation and SLA compliance."
Cross-Platform Expo / React Native Mobile App with Adaptive Coaching & In-App Entitlements
A shipped cross-platform mobile application built for the RevenueCat Shipaton 2026 on Expo and React Native, featuring an adaptive habit intelligence engine, streak-shielding burnout defense, and native RevenueCat subscription paywalls.
Key Engineering Highlights:
- Adaptive coaching engine analyzing completion velocity, time-of-day friction, and streak risk to recommend habit stacking and burnout recovery.
- Native RevenueCat SDK (react-native-purchases) integration gating pro_access entitlements across Monthly, Annual, and Lifetime paywall offerings.
- Abstracted purchase provider architecture toggling seamlessly between live App Store / Play Store billing and a deterministic simulator store for CI/CD.
- Offline-first AsyncStorage state persistence with 36/36 unit tests passing across habit logic, entitlement gating, and coaching heuristics.
ExpoReact NativeRevenueCat SDKTypeScriptAsyncStorageMobile PaywallsVitest
"Shipped HabitFlow Pro for RevenueCat Shipaton 2026, a cross-platform Expo and React Native mobile app with an adaptive habit coaching engine, offline-first AsyncStorage sync, RevenueCat pro_access entitlement paywalls, and 36/36 passing unit tests."
Interactive Architecture Canvas, Time-Synced Transcript Search, SM-2 Flashcards & Obsidian Exporter
An interactive full-stack distributed systems curriculum, spaced-repetition trainer, and architectural interview simulator built with Next.js 14 and FastAPI 2.0 Async.
Key Engineering Highlights:
- Curated system design catalog covering Rate Limiters, Raft/Paxos Consensus, Consistent Hashing, Sharding, Kafka Queues, and API Gateways with Mermaid/Cytoscape canvases.
- SuperMemo-2 (SM-2) spaced repetition engine computing dynamic easiness factors (EF) and optimal review intervals from user confidence ratings (0-5).
- Time-synchronized video transcript parser and full-text chapter search engine powered by asynchronous SQLAlchemy and PostgreSQL/SQLite.
- 1-click Obsidian Vault ZIP exporter generating cross-linked Markdown notes with YAML frontmatter, plus a 45-minute FAANG mock interview simulator.
Next.js 14TypeScriptFastAPISQLAlchemy AsyncPostgreSQL / SQLiteMermaid.jsSM-2 AlgorithmObsidian Export
"Architected System Design Master Vault, a full-stack distributed systems platform in Next.js 14 and FastAPI featuring time-synced transcript search, an SM-2 spaced repetition engine, interactive Mermaid canvases, and 1-click Obsidian vault export."
10 End-to-End Instructional Packages on Autonomous TDD Loops, MCP Servers, Custom Skills & AGENTS.md
A comprehensive suite of 10 production-ready instructional engineering packages and working codebases demonstrating advanced software engineering workflows with the OpenAI Codex CLI and VS Code Extension.
Key Engineering Highlights:
- 10 complete hands-on modules covering Autonomous TDD Loops, Custom Reusable Skills (.codex/skills), Visual Multi-File Refactoring, and Model Context Protocol (MCP) database inspection.
- Deterministic agent governance patterns using hierarchical AGENTS.md rules to enforce package whitelists, error hierarchies, and pytest execution gates.
- Automated AI Code Review workflows ('Who Reviews the AI?') catching SQL injection, race conditions, and subtle regressions in AI-generated pull requests.
- Includes working Python/FastAPI demo repositories, pytest verification suites, slide decks, and second-by-second narration scripts.
OpenAI CodexModel Context ProtocolAGENTS.mdPythonFastAPIPytestVS Code Extension
"Authored a 10-module OpenAI Codex Engineering Curriculum covering autonomous TDD loops, custom .codex/skills, live MCP database inspection, AGENTS.md governance invariants, and automated AI PR security code reviews."
Spec-Driven Omnichannel Creative Synthesis, AdCP Media Buying, A2A Rate Negotiation & 4-Gate Evals
A production-grade agentic advertising platform connecting omnichannel creative generation (Display, Video, CTV, DOOH), Ad Context Protocol (AdCP) / OpenRTB media discovery, 3-round A2A rate negotiation, HITL spend activation, and a 4-Gate automated evaluation harness.
Key Engineering Highlights:
- Stateful LangGraph v1.x orchestration with parallel
Send() creative fan-out, Command routing, time-travel checkpoint forking, and interrupt() HITL approval gates.
- Official AdCP over MCP tasks (
get_adcp_capabilities, get_products, sync_creatives, create_media_buy) with penny-perfect integer budget conservation and 3-round A2A CPM rate negotiation.
- 5-Point Multimodal Creative QA enforcing IAB dimensions, CTV 10% overscan safe-zones, DOOH ≤8-word legibility, WCAG 2.2 AAA contrast, and C2PA provenance manifests.
- Agent Harness enforcing
HMAC-SHA256 JWT live-spend guards, SHA-256 idempotency, FinOps circuit breakers, Saga rollback (cancel_media_buy), and GDPR cascading right-to-erasure.
Python 3.13FastAPILangGraph v1.xAdCP / OpenRTBA2A ProtocolHybrid RAG (RRF)VAST 4.2 / C2PASQLite WAL4-Gate Evals
"Architected AdAiz, an enterprise agentic advertising platform uniting LangGraph v1.x parallel Send() fan-out, AdCP over MCP media buying, 3-round A2A rate negotiation, HMAC-SHA256 live-spend HITL gates, 5-point multimodal creative QA, and a 4-Gate CI evaluation suite."
Idempotent Commercial Quote Engine, MCP Rate Server, Azure APIM Gateway & Foundry Evals
A dual-plane enterprise system pairing a Teams Copilot Studio Delivery Desk (turning scope changes into validated, idempotent commercial quotes) with an Azure AI Foundry evaluation and APIM governance Control Plane.
Key Engineering Highlights:
- Single-source-of-truth commercial rate card (
ratecard/policy.py) shared identically across the MCP tool server (get_rate, create_quote), quote math, and evaluation suite.
- Zero-privilege boundary design: the Teams agent has no send-mail tool; Power Platform cloud flows own human approval and vendor rejection paths.
- Azure API Management (APIM) gateway policy enforcing managed identity authentication, model routing, token rate limits, and
x-harbor-gateway: apim header validation.
- 25-case offline Foundry evaluation harness and full Bicep IaC (
infra/main.bicep) with federated GitHub Actions deployment.
PythonAzure AI FoundryCopilot StudioMCP ServerAzure APIMBicep IaCPower PlatformPytest
"Engineered Harbor Field Services, pairing a Microsoft Teams Copilot Studio Delivery Desk and idempotent MCP quote engine with an Azure AI Foundry 25-case evaluation harness, APIM token-limiting gateway policies, and Bicep IaC."
Multi-Project A/B Testbed Benchmarking 24 Agent Skill Packs Across the Full SDLC
A comprehensive 8-domain evaluation laboratory for benchmarking, comparing, and regression-testing AI coding agent skills across UI/UX, Full-Stack SaaS, Autonomous Agents, Cloud IaC, OWASP Security, Testing/QA, Databases, and DevOps CI/CD.
Key Engineering Highlights:
- 8 isolated evaluation sandboxes (
eval-ui-ux, eval-fullstack, eval-agentic-ai, eval-core-engineering, eval-security-audit, eval-testing-qa, eval-database-performance, eval-devops-cicd).
- Side-by-side Git branch A/B evaluation workflow comparing 24 candidate skill packs under identical
EVAL_BENCHMARK.md specification contracts.
- Tests hard engineering invariants: DTCG design tokens, WCAG 2.2 AAA contrast, ReAct prompt injection fences, IDOR prevention, Red-Green TDD with
fast-check, and zero-downtime expand-contract DB migrations.
Agent EvalsTypeScriptPythonVitest / Playwrightfast-checkOWASP SecurityRedis MutexTerraform IaC
"Created Skills Evaluation Lab, an 8-domain multi-project benchmark suite evaluating 24 AI coding agent skill packs across WCAG 2.2 UI/UX, OWASP IDOR security, ReAct sandbox isolation, property-based TDD, and zero-downtime database migrations."
Latent Representation Honesty Direction Vector & Truthfulness Decoding Guardrail
Built for the IBM Research / NeurIPS 2026 Steerability Challenge, HonestSteer implements a standardized .spipe activation steering pipeline that suppresses dishonest LLM assertions without degrading general reasoning capabilities.
Key Engineering Highlights:
- Injects calibrated honesty direction vectors directly into intermediate transformer residual stream activations during forward passes.
- Preserves MMLU/GSM8K general reasoning capability by bounding activation norm perturbations via cosine-similarity gating.
- Exports fully compliant
.spipe steering pipeline artifacts with 100% passing verification tests.
Python 3.11+Activation Steering.spipe StandardNeurIPS 2026IBM ResearchLLM AlignmentPytest
"Developed HonestSteer for the NeurIPS 2026 / IBM Steerability Challenge, implementing a standardized .spipe residual-stream activation steering pipeline that suppresses dishonest LLM completions without degrading benchmark reasoning."
Point-in-Time Lagged Alpha Generator, Dollar-Neutral Allocator & G0-G3 Audit Guard
Built for the NeurIPS 2026 Agenthon (SQA & CEWIT), VeriQuant AI is an autonomous, leak-proof quantitative alpha generation and portfolio optimization agent engineered to pass strict cheat-resistant evaluation gates (G0–G3).
Key Engineering Highlights:
- Strictly lagged point-in-time feature pipeline (
t-1) preventing lookahead bias and forward-return leakage under adversarial temporal truncation tests.
- Cross-sectional rank normalizer paired with a dollar-neutral, turnover-constrained mean-variance portfolio optimizer.
- Automated G0–G3 gate auditor validating schema integrity, zero future leakage, turnover caps, and deterministic Docker reproducibility.
Python 3.11+Quant AlphaNeurIPS 2026 AgenthonMean-Variance OptLeakage ControlDockerPytest
"Engineered VeriQuant AI for the NeurIPS 2026 Quantitative Agenthon, combining a strictly lagged (t-1) point-in-time alpha generator, cross-sectional rank normalizer, and dollar-neutral portfolio allocator passing all G0–G3 anti-leakage audit gates."
Planner-Executor-Verifier Multi-Agent Swarm for Lablab.ai AI Genesis 2026
An autonomous multi-agent enterprise operational workflow orchestrator built for AI Genesis 2026 (Lablab.ai), combining structured DAG task decomposition, parallel tool execution, and a deterministic ground-truth verification gate.
Key Engineering Highlights:
- Three-stage Planner → Tool Execution Swarm → Deterministic Verifier Gate pipeline preventing unverified agent outputs from reaching downstream systems.
- Strict TypeScript state machine with deterministic schema validation and interactive real-time execution dashboard.
TypeScript (Strict)Multi-Agent SwarmVerifier GateLablab.aiAI Genesis 2026
"Built Genesis Ops Mesh for AI Genesis 2026, a strict TypeScript multi-agent workflow orchestrator pairing DAG task decomposition and parallel tool execution with a deterministic ground-truth verification gate."
No projects match your query
Try clearing your search query or selecting a different category tab.