NexusRec AI — Multimodal Visual-Semantic Retail & Industrial Discovery Engine

Conversational & Visual Multimodal Discovery Console

Select a buyer persona or enter natural language + visual reference constraints to inspect the live 4-stage recommendation funnel.

Quick Scenarios:

Active Buyer Persona (UserTower Context)

Recent session: RET-005 (B3 Crampon Boots), RET-013 (12-Pt Crampons), RET-019 (Merino Baselayer)

32-D OpenCV Visual Reference Descriptor

Extracts 12-bin HSV color + 10-bin Sobel contour + 10-bin Gabor surface energy.

MMR λ (Rel/Div) 0.72

Lower λ = wider categories

Bandit Explore 0.12

Cold-start Beta(α,β) boost

Executing multimodal Two-Tower + BM25 RRF retrieval...

Pipeline Latency 3.8 ms total

SASRec Causal Self-Attention

Live Clickstream Session Trajectory (`last_5_clicks` in NoSQL Online Store)

Anchor: $229.67 Session CTR: 33.3% Embedding Drift: 0.0000

Predicted Next Categories: Climbing Hardware → Alpine Shells

STAGE 1 • CANDIDATE GEN 44 → 21 items

Qdrant Filter + Two-Tower 64-D

Hard payload filter (in_stock, max_price, domain) + L2 User/Item Tower cosine dot product.

STAGE 2 • HYBRID RRF FUSION Top 18 fused

Dense Cosine + BM25 (k = 60)

Reciprocal Rank Fusion merges 64-D multimodal vector ranks with sparse SPLADE/BM25 lexical ranks.

STAGE 3 • DEEP RERANKING GBDT + Bandit

Cross-Encoder + Thompson Bandit

Scores 7 pairwise feature crosses (compatibility, budget headroom, gross margin) + Beta(α,β) exploration.

STAGE 4 • MMR DIVERSITY 6 Final Items

Maximal Marginal Relevance

Penalizes redundant category & visual clones so buyers receive a complete, high-margin kit.

Ranked Multimodal Recommendations (Click Bandit Reward to Update Posterior Live)

▶ Inspect Live Qdrant Hybrid Payload Filter, Visual Descriptor & 64-D UserTower Tensor ▸

Qdrant Payload Filter & Latency Breakdown

UserTower 64-D L2-Normalized Embedding & Visual Query Signature

Offline & Online RecSys Evaluation Suite (K = 10)

Deterministic comparison across 4 enterprise buyer cohorts: Baseline Popularity vs. Pure Collaborative Filtering vs. NexusRec Two-Tower + MMR.

Architecture	NDCG@10	Recall@10	Precision@10	MRR	IPS	SNIPS (Debiased)	Coverage	Diversity (ILD)	Avg Margin

InfoNCE Contrastive Loss (τ = 0.08)

1.9840 → 0.4120

In-batch sampled softmax aligns positive (user, item) pairs on the 64-D unit hypersphere.

Positive vs. Negative Cosine Margin

+0.6920 separation

Mean positive pair cosine: 0.7840 • In-batch negative cosine: 0.0920

Tensor Execution Backend

64-D L2 Hypersphere

32-D OpenCV HSV/Sobel/Gabor Visual + 32-D Semantic/Commercial Taxonomy.

Per-Cohort NexusRec Performance Breakdown

Hybrid SQL + NoSQL Feature Store & Cross-Framework PyTorch / TensorFlow / ONNX Parity

Point-in-time SQL offline feature joins (as_of_timestamp leakage guard), Redis-compatible NoSQL session documents, and TensorFlow SavedModel / TFLite INT8 / ONNX Runtime numerical equivalence.

SQLite3 Offline Feature Store

12 Interaction Rows

Point-in-Time Join Audit: PASSED_ZERO_FUTURE_LEAKAGE (Blocked future rows > t_cutoff)

NoSQL Online Session Store

98.5% Cache Hit Rate

Sub-ms TTL document store serving last_5_clicks, session_price_anchor & real_time_ctr.

PyTorch ↔ TF ↔ ONNX Parity

Max Diff: 0.00e+00 (< 1e-5)

TFLite INT8 Cosine Fidelity: 0.99989 • 4.0x Memory Footprint Compression

▶ Inspect Live Feature Store Point-in-Time Audit & Cross-Framework Tensor Export JSON JSON

Production Cloud Deployment Blueprints (AWS & GCP Scale-to-Zero)

Expand any infrastructure manifest below to inspect the production Terraform, Google Cloud Run GPU/CPU serving configuration, or Docker multi-stage build.

▶ Inspect Google Cloud Run Knative Manifest (deploy/gcp/service.yaml — Scale-to-Zero NVIDIA L4 / CPU) YAML

apiVersion: serving.knative.dev/v1
kind: Service
metadata:
  name: nexusrec-ai-discovery
  annotations:
    run.googleapis.com/ingress: all
    run.googleapis.com/launch-stage: BETA
spec:
  template:
    metadata:
      annotations:
        autoscaling.knative.dev/minScale: "0"
        autoscaling.knative.dev/maxScale: "12"
        run.googleapis.com/cpu-throttling: "false"
        run.googleapis.com/startup-cpu-boost: "true"
    spec:
      containerConcurrency: 80
      timeoutSeconds: 60
      containers:
        - image: us-central1-docker.pkg.dev/PROJECT_ID/nexusrec-repo/nexusrec-ai:latest
          ports:
            - containerPort: 8000
          resources:
            limits:
              cpu: "4000m"
              memory: "8Gi"
          startupProbe:
            httpGet:
              path: /api/v1/health
              port: 8000

▶ Inspect AWS Terraform Infrastructure (deploy/aws/main.tf — ECS Fargate + SageMaker Serverless + ElastiCache Redis) HCL / Terraform

resource "aws_sagemaker_model" "two_tower_onnx" {
  name               = "nexusrec-two-tower-64d"
  execution_role_arn = aws_iam_role.sagemaker_exec.arn
  primary_container {
    image          = "${aws_ecr_repository.nexusrec.repository_url}:latest"
    model_data_url = "s3://${aws_s3_bucket.model_registry.id}/models/two_tower_64d.tar.gz"
  }
}

resource "aws_sagemaker_endpoint_configuration" "serverless_cfg" {
  name = "nexusrec-serverless-cfg"
  production_variants {
    variant_name = "TwoTowerPrimary"
    model_name   = aws_sagemaker_model.two_tower_onnx.name
    serverless_config {
      memory_size_in_mb = 4096
      max_concurrency   = 32
    }
  }
}