Open to Senior Applied AI · Forward-Deployed Engineering · AI Solutions Architecture roles
Production · Live system · Engineering · Stack · Principles · Repositories
I design, build and operate production AI systems with end-to-end technical ownership — from computer-vision models and GPU inference to LLM agents, business integrations, backend/data layers and product interfaces.
My bias is toward measurable impact, honest evaluation and systems that stay reliable after the demo is over.
|
Retail outlets ~1,500 active monthly |
Orders / month 98% via platform |
Detection F1 on unseen shelf images |
Brand precision confirmed end-to-end evaluation |
| 🧠 20+ typed tools production agent layer |
⚡ NVIDIA H200 self-hosted AI inference |
🤖 Self-hosted Qwen Qwen2.5-VL 72B · Qwen3.6 35B served with vLLM |
🎯 73.1% SKU precision end-to-end retrieval + fusion |
🔒 Most production code is private because it runs inside commercial systems with proprietary data and business integrations. The public repositories below contain sanitized case studies, architecture, metrics, evaluation methodology and runnable examples.
Real shelf photo through the production CV pipeline — detections, SKU labels, price-tag reads and explicit Unknown when evidence is insufficient.
flowchart LR
A[Retail shelf photo] --> B[GroundingDINO detection]
B --> C[Product crops]
C --> D[Qwen2.5-VL OCR / package evidence]
D --> E[Qwen3-Embedding-8B]
E --> F[Qdrant dense retrieval]
C --> G[DINOv2 visual k-NN]
C --> H[ArcFace metric retrieval]
F --> I[Deterministic fusion + guardrails]
G --> I
H --> I
D --> I
I --> J{Enough evidence?}
J -->|Yes| K[SKU / brand match]
J -->|No| L[Unknown / abstain]
Production principle: confidence is not correctness. A real AI system should be able to abstain, run in shadow before promotion, be evaluated on real populations and roll back cleanly.
|
Production case studies with architecture, metrics, engineering decisions and evaluation methodology.
Includes:
|
Production-engineered retrieval + computer-vision merchandising pipeline with explicit uncertainty.
Evidence:
|
Multi-source marketing intelligence platform connecting social, website, search and campaign analytics.
Measured window:
|
Computer Vision · OCR · Vision-Language Models · Dense Retrieval · Retrieval-Augmented Systems · Vector Search · LLM Agents · Tool Calling · MCP · GPU Inference · Evaluation · Observability · Marketing Intelligence
| 🎯 Evaluate first Golden-set regression, replay, acceptance gates and population-level validation before promotion. |
🛡 Safe rollout Shadow → measure → gate → active, with explicit rollback paths. |
⚙️ Deterministic where possible LLMs augment reliable systems; they do not replace reliable logic without reason. |
🏗 End-to-end ownership Model layer, backend, data, integrations, frontend, deployment and operations. |
