Building the Future of AI-Powered Platforms & Agentic Systems
Hi, I'm Akmal Khan. I am a Senior Full Stack AI Engineer and systems architect with 20+ years of software engineering experience, including 3+ years dedicated to architecting and shipping production AI platforms. I engineer end-to-end intelligent systems across the full stack—from TypeScript/React, Next.js, and React Native applications to high-throughput Node.js (Hono, Express) and Python backends (FastAPI, Django). I own every layer of production AI: multi-agent autonomous workflows (LangGraph, CrewAI, MCP), Claude/OpenAI tool use and streaming pipelines, pgvector semantic search, and resilient distributed asynchronous job processing (Redis, BullMQ, Celery, Kafka).
Core Focus
About Me
I focus on building production-grade, scale-tested, and secure AI-powered platforms. My experience covers full-stack web and mobile architectures, distributed systems, and integration of Large Language Models (LLMs) into autonomous workflows using tools like LangGraph, CrewAI, LlamaIndex, and custom Model Context Protocol (MCP) servers.
I treat AI pipelines as engineered systems: instrumenting token economics, latency, and extraction accuracy. Whether it's implementing ACID transactions to prevent financial data corruption, scaling background jobs using Redis and BullMQ, or designing pgvector embeddings for RAG searches, I ensure that design choices prioritize correctness, performance, and security.
Core Principles
AI & RAG Engineering
Expertise in Claude API, tool use/function calling, pgvector indexing, and hybrid search RAG evals.
Robust Async Systems
Designing decoupled background processing with Redis, BullMQ, and Celery job queues.
Security & Observability
Enforcing strict RBAC, data isolation, automated regression test runs, and token cost dashboards.
Technical Stack
Agentic AI, Voice & RAG
- LangGraph v1.x, CrewAI & A2A
- FastMCP, AdCP & Claude Tools
- LiveKit SIP, WebRTC & Silero VAD
- pgvector, Qdrant & Hybrid RRF
- 4-Gate Evals & Activation Steering
- Gemini Live, Deepgram & Cartesia
- YOLO26, Detectron2, SAM 2 & OpenCV
- PyTorch Two-Tower RecSys & ONNX/TensorRT
Systems, Backend & Async
- Python 3.13+ (FastAPI, Django)
- Go 1.25+ & Rust 2024 (SIMD DSP)
- Node.js (TypeScript, Hono, Express)
- Redis, BullMQ, Celery & Kafka
- ISO 8583 JIT & Double-Entry Ledgers
- DuckDB Parquet & Airflow 3.0 ETL
Frontend, Audio & Mobile
- React 19 & Next.js 16 (App Router)
- Angular 21 Signals & TypeScript
- AudioWorklet & SharedArrayBuffer
- React Native & Expo (RevenueCat)
- Accessible UI & Tailwind CSS
- Vitest, Playwright & Pytest TDD
DB, Cloud & Governance
- PostgreSQL 17 (JSONB, HNSW Tuning)
- AWS (Bedrock, ECS, Chime SIP, S3)
- Azure AI Foundry, APIM & Bicep IaC
- GCP Cloud Run & Vultr CloudOps
- Docker & Eval-Gated GitHub Actions
- SPIFFE/SPIRE Zero-Trust & RBAC
Education & Certs
PhD in Computer Science
Seoul National University, South Korea · 2014
MS in Computer Engineering
CASE, Islamabad · 2008
Certifications
- LangChain & Vector DBs (Activeloop) · 2024
- GCP Professional Developer · 2024
- AWS Cloud Solutions Architect · 2023
- Deep Learning Specialization · 2022
Experience
Senior AI & Full Stack Engineer
Shipped AI platforms with Claude API, pgvector RAG, and async job queues for enterprise clients. Reduced LLM costs by 30-40%.
Senior Full Stack Engineer & Architect
Architected and delivered 10+ remote SaaS products across fintech, e-commerce, and workforce management.
Software Engineer & Lecturer
Developed university web systems and lectured in software engineering and machine learning at Islamia University of Bahawalpur.
Engineering AdAiz: Designing an Agentic Advertising Platform with AdCP over MCP, 3-Round A2A Rate Negotiation, and LangGraph HITL Spend Gates
How to architect a deterministic autonomous media-buying platform: bridging Ad Context Protocol (AdCP) over MCP, 3-round Agent-to-Agent (A2A) CPM rate negotiation, HMAC-SHA256 live-spend circuit breakers, and 5-point omnichannel creative QA.
Tech Blog
I write deep-dives on system design, AI orchestration, and database scaling.
Visit Writing HubFeatured Flagship Projects
Studio OS: AI Film Studio Operating System
An AI-powered film studio operating system and multi-agent network automating video ingestion, screenplay analysis, casting performance, and auto-dubbing translation.
- 7 specialized agents coordinate video ingestion, pre-production breakdown, and continuous takes analysis using Gemini 2.5 Pro vision.
- Offloads audition recording analysis (pitch, tempo, emotional tracking) to casting agents.
- Unified backend API Gateway (FastAPI) serving a Next.js web panel and an Expo mobile app.
SpecForge: AI Requirements Extraction
An enterprise AI requirements platform that parses PDF/DOCX PRDs into structured epics, user stories, and Gherkin acceptance criteria.
- Two-pass Claude 3.5 extraction pipeline with JSON schema validation.
- Semantic story deduplication via pgvector embeddings.
- Live token economics and cost-tracking dashboard.
Ops MCP Suite: Production Toolkit
A production-grade collection of 6 Model Context Protocol (MCP) servers (DevOps, MLOps, AgentOps, DB, GitHub, Secrets) served via a unified FastAPI gateway.
- Provides Supervisor networks with serverless database replication and evaluations.
- Containerized and deployable to Google Cloud Run.
Agent Fleet Orchestrator: Multi-Agent Supervisor
A distributed agent coordination engine implementing supervisor-worker topologies for dynamic task delegation, state checkpoints, and human-in-the-loop approvals.
- Hierarchical LangGraph state graph with persistent memory checkpoints.
- Asynchronous parallel task dispatching to distributed worker nodes.
- Real-time telemetry and execution trajectory audit logs.
LeaseLogic AI: Commercial Lease Extraction Engine
An AI-powered commercial real estate lease intelligence platform that extracts financial clauses, rent escalations, and termination liabilities from 100+ page contracts.
- Claude 3.5 Sonnet tool use and JSON schema enforcement for guaranteed tabular clause extraction.
- Hybrid dense vector + BM25 keyword search across multi-tenant pgvector indices for audit citation.
- Real-time financial calculation engine for Net Effective Rent and NPV modeling.
ClaimPilot: AI Insurance Claims Adjudication
An AI-driven claims triage and policy verification engine that automates first-notice-of-loss (FNOL) review, document verification, and fraud detection heuristics.
- Automated extraction of policy bounds, deductible schedules, and claim validation.
- Multi-modal damage evidence analysis and cross-verification against historical claims.
- Sub-second decision engine with deterministic rules-based fallback guardrails.
IntakeRx: Clinical Triage & EHR Automated Ingestion
A HIPAA-compliant clinical intake platform that converts unstructured patient voice interviews and handwritten forms into standardized FHIR/HL7 EHR records.
- Offline-capable React Native mobile client for decentralized clinic patient intake.
- Automated clinical SOAP note formatting with dual-pass safety and allergen conflict verification.
- Immutable database audit log triggers and AES-256 field-level encryption for PHI data.
MedEdge: Hybrid Clinical Assistant & 2G SMS
A cross-platform, local-first clinical decision assistant designed for resource-constrained clinics. Operates offline with Gemma via Ollama or online with cloud Gemini models.
- Consultation pipeline: transcription → SOAP Clinical Note formatting → Triage draft.
- 2G SMS fallback bridge via Africa's Talking API for messaging feature-phone patients.
- Admin web panel toggles local offline LLM providers and monitors local server health.
Secure Healthcare Audit Vault
A high-security, HIPAA-compliant patient record repository featuring document QA search using vector embeddings and PostgreSQL.
- Vector Search RAG using `pgvector` inside Postgres (with HNSW index optimization).
- Enforced database-level immutable audit logging through PG triggers to prevent tamper.
- Temporary AWS S3 Signed URLs for zero direct-link exposure.
High-Throughput Async Payroll Engine
A high-performance async payroll batch system engineered to handle large financial calculations and payslip exports asynchronously in celery background tasks.
- Offloads CPU-bound PDF rendering to multi-process Celery workers.
- Guaranteed financial correctness using Python's `Decimal` type instead of floating points.
- ACID transactional integrity via `@transaction.atomic` with automatic rollback on errors.
AI Care Agency Rota Manager
An AI-powered care scheduling platform that allows agency coordinators to assign shifts and check compliance using natural language prompts. Offloads scheduling validation to an agentic tool loop.
- Interpret commands like *"Assign Alice to Bob on Monday"* via OpenAI function calling.
- Real-time certification check and automatic double-booking prevention via Postgres constraints.
- Dockerized environment with PostgreSQL database, Express backend, and Next.js UI.
Procurement Intelligence Multi-Agent System
An enterprise procurement intelligence suite with LangGraph orchestrators, Vanna AI Text-to-SQL, AWS Bedrock Guardrails, and Pinecone vector search for purchase compliance.
- Multi-agent orchestration validating enterprise RFP compliance and contract bounds.
- Natural language SQL generation using Vanna AI over enterprise vendor databases.
- AWS Bedrock Guardrails ensuring strict deterministic policy enforcement.
Agentic AI Applications Portfolio (17 Microservices Monorepo)
A unified monorepo showcasing a suite of 17 business-oriented AI agent microservices running parallel RAG, voice automation, and visual QA tests.
- Autonomous Coder running inside isolated E2B code execution sandboxes.
- Event-driven Fintech Fraud detection checking transactional patterns over Apache Kafka.
- Unified React dashboard to control, execute, and monitor real-time telemetry from all 17 agents.
ScholarAssist: Research Paper Assistant
An academic research and literature synthesis tool that ingests scientific PDF papers, builds cross-citation graphs, and grounds literature review drafting.
- Dense vector retrieval with precise paragraph citation and hallucination minimization.
- Comparative matrix generation across multiple published methodology papers.
- BibTeX reference export and automated executive summaries.
QuestionPaperAI: Assessment Platform
An AI-powered intelligent assessment platform that streamlines test paper generation and student management through adaptive templates and question bank verification.
- Generates unique, randomized assessment papers and grading rubrics for individual students using OpenAI.
- Maintains a secure Question Bank with strict verification pipelines prior to inclusion.
- Features document exporting capabilities for generating beautifully formatted DOCX and PDF documents.
Kaggle Competition Master: ML Pipeline Dashboard
A machine learning competition engineering workbench with automated feature engineering, cross-validation tracking, and model ensembling pipelines.
- Automated LightGBM, XGBoost, and PyTorch deep neural net stacking.
- Optuna Bayesian hyperparameter optimization and cross-validation metrics tracking.
- Interactive evaluation dashboard for feature importance and confusion matrix inspection.
Python Interview Prep Suite
A comprehensive preparation suite of 16 structured Jupyter Notebooks covering advanced paradigms, memory management, and cross-cloud systems engineering.
- Exhaustive coverage of Object-Oriented Design (OOD) principles, algorithms, and concurrency patterns.
- Includes detailed technical mappings of service configurations between AWS, GCP, and Azure.
- Demonstrates performance-tuning strategies for core Python architectures and memory management.
SawtOS: Carrier-Grade Real-Time Voice & Telephony Engine
An enterprise conversational voice and SIP telephony platform for US and Gulf markets with a Dual-Engine Voice Router (Native S2S + Cascaded), sub-4ms pgvector RAG, and HIPAA clinical ETL.
- 285ms E2E latency across Silero VAD v5 (1.2ms), Deepgram Nova-3, Groq Llama 3.3 70B, and Cartesia Sonic 3.5.
- LiveKit SIP trunking (+1 US, +966 KSA, +971 UAE) with instant <35ms barge-in buffer flush.
- Post-call LLM-as-a-Judge QA scoring via Langfuse and Apache Airflow 3.0 HIPAA Safe Harbor de-identification.
TrafficSentinel Suite: Ad-Tech IVT Investigator & MCP Platform
An enterprise ad-tech trust & safety monorepo combining an interactive clarifying investigator agent, an in-process DuckDB bidstream FastMCP server, and a 50-scenario CI evaluation harness.
- Programmatic AmbiguityGate pausing underspecified queries with 100% Clarification Trigger Rate (CTR).
- High-throughput DuckDB telemetry MCP server over 500k+ Parquet bidstream records (<800-token budget).
- Angular 21 Signal-based reactive dashboard and regression diff viewer catching prompt drift in CI/CD.
VoicePilot AI: Full-Duplex Enterprise Voice Assistant
A full-duplex conversational voice AI platform engineered for sub-300ms roundtrip latency, instant barge-in interruption cutoff (<20ms), and live waterfall latency telemetry.
- 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 reactive canvas frequency orb HUD.
- Streaming STT (Deepgram Nova-3), Groq Llama 3.3 70B (<90ms TTFT), and Kokoro-82M / Cartesia TTS.
OmniCast Studio: Multimodal Research, Graph RAG & Broadcast
An open-source multimodal research and broadcast platform transforming dense papers and repos into interactive GraphDB mind maps, EBU R128 dual-host podcasts, Remotion videos, and live voice chat.
- Interactive @xyflow/react + Memgraph/Kùzu concept mind maps with 1-click source citation jumping.
- Dual-host podcasts mastered to -16 LUFS via Rust SIMD DSP alongside 16:9 & 9:16 Remotion video rendering.
- Full-duplex 'Interrupt & Ask' voice mode over Gemini Multimodal Live WebSockets.
OmniCard & FX: JIT Card Authorization Gateway
A spec-driven virtual card issuing, sub-10ms ISO 8583 Just-In-Time (JIT) authorization gateway, dynamic multi-currency FX engine, and Maker-Checker governance platform in Go.
- Evaluates card status, MCC rules, velocity caps, and atomic balance holds in under 10ms.
- Guaranteed 60-second FX quote rate locks with basis-point spread buffers and dual-control governance.
ReconMesh: EOD 3-Way Financial Reconciliation
A high-throughput Go reconciliation engine verifying data integrity across internal double-entry ledgers, CAMT.053/BAI2 bank statements, and payment processor webhooks.
- Reconciles 10,000 multi-party transactions in <50ms across 3 matching tiers (Exact, Windowed, Batch).
- Auto-generates double-entry suspense ledger entries for unmatched breaks and FX slippage.
IntentGuard: AI Coding Control Plane
An inline governance control plane, blast-radius circuit breaker, and 1-click scope-spill rollback suite for Antigravity, Claude Code, and OpenAI Codex.
- Halts multi-hotspot thrashing and unexpected file truncations mid-flight.
- Ships as a VS Code VCR cockpit, headless CLI (
ig), and GitHub Action PR gatekeeper.
AdAiz: Enterprise Agentic Advertising Platform
A spec-driven autonomous advertising platform uniting omnichannel creative synthesis (Display, Video, CTV, DOOH), AdCP / OpenRTB media discovery, 3-round A2A rate negotiation, and 4-Gate automated evaluations.
- Stateful LangGraph v1.x orchestration with parallel
Send()creative fan-out, time-travel checkpoint forking, andinterrupt()HITL live-spend gates. - HMAC-SHA256 JWT live-spend guardrails, penny-perfect integer budget conservation, and Saga compensating rollback (
cancel_media_buy). - 5-point multimodal creative QA (CTV 10% overscan safe-zone, DOOH ≤8 words, WCAG 2.2 AAA contrast, and C2PA provenance).
Harbor Field Services: Quote & Control Plane
An enterprise Copilot Studio Delivery Desk and Azure AI Foundry Control Plane turning Teams scope-change requests into validated, idempotent commercial quotes.
- Single-source-of-truth rate card engine exposed via MCP (
get_rate,create_quote) behind Azure APIM gateway policies. - 25-case Foundry offline evaluation harness, Bicep IaC, and Power Platform human approval cloud flows.
Skills Evaluation Lab: 8-Domain AI Coding Agent Benchmark Suite
A multi-project evaluation testbed for A/B benchmarking AI coding agent skills across UI/UX, Full-Stack SaaS, Autonomous Agents, Cloud IaC, OWASP Security, Red-Green TDD, Database Indexing, and Hardened CI/CD.
- 8 isolated domain sandboxes (
eval-ui-uxthrougheval-devops-cicd) comparing 24 candidate skill packs under identical specification constraints. - Deterministic pass/fail verification rubrics testing DTCG design tokens, IDOR prevention, zero-downtime expand-contract migrations, and property-based testing.
AegisVision AI: 3-Tier Industrial Visual Inspection & CAPA Copilot
A 3-tier cascaded computer vision and MLOps platform for high-speed manufacturing (PCB SMT, 4680 EV Battery Laser Welds, Pharma Blisters, Wafers) and automated retail planogram auditing.
- Sub-12ms Tier-1 OpenCV 3×3 planar homography + CLAHE, YOLO26 NMS-Free detection, and Detectron2 Mask R-CNN pixel metrology (μm² / mm²).
- IsolationForest OOD gate routing uncertain/novel anomalies to Meta SAM 2 zero-shot mask refinement, achieving 126.6 FPS (5.21× speedup) on TensorRT 10 / ONNX INT8.
- Hybrid Dense+BM25 RRF RAG over IPC-A-610H, IATF 16949, FDA 21 CFR, and GS1 SOPs with automated YOLO/COCO active-learning pseudo-label distillation.
NexusRec AI: Multimodal Visual-Semantic Recommender & Commerce Engine
A 4-stage personalization, visual search, and conversational commerce engine spanning Retail Technical Apparel and Industrial MRO Spare Parts.
- 64-D L2-normalized PyTorch Two-Tower neural retriever (OpenCV HSV/Sobel/Gabor visual descriptors + semantic embeddings) trained with In-Batch Negative InfoNCE loss.
- Qdrant Universal Hybrid Search (Dense Cosine + Sparse BM25 RRF) with hard inventory/budget payload pushdown filters in <8ms.
- Scikit-learn Gradient Boosted Cross-Encoder reranking + Thompson Sampling contextual bandits + Maximal Marginal Relevance (MMR) diversity (0.928 NDCG@10, 0.958 MRR).
Let's build something remarkable
I am currently open to new career opportunities and consulting projects. If you are looking for a full-stack engineer who understands AI Agent architectures and backend scalability, let's connect!