Flagship project
Atlas
An agentic retrieval system that decomposes a question, routes across five source types and returns an answer with paragraph-level citations.
- 01
- Python
- 02
- LangGraph
- 03
- pgvector
- 04
- Cohere Rerank
- 05
- FastAPI
- 06
- Redis
AI / ML Engineer — Remote
NaveedAshfaq
I build language and vision systems that survive production traffic.
- LLM systems
- RAG pipelines
- AI agents
- Computer vision
- MLOps
Selected work
Systems, not demos.
Each of these went to production — with evaluation harnesses, latency budgets and a real user on the other end.
Flagship project
Sentinel
Real-time defect detection running at 60 FPS on a Jetson Orin Nano — quantised, distilled and deployed to 140 production lines.
- 01
- PyTorch
- 02
- TensorRT
- 03
- ONNX
- 04
- OpenCV
- 05
- Jetson
- 06
- MLflow
Flagship project
Ledger
Structured extraction from 40-page financial PDFs with schema validation and a confidence score per field, replacing a nine-person review team.
- 01
- Python
- 02
- Pydantic
- 03
- Claude
- 04
- Tesseract
- 05
- Postgres
- 06
- Temporal
The engineer
Research-grade models,
production-grade systems.
- 0
- Systems in production
- 0+
- Models fine-tuned
- 0+
- Eval suites written
- 0
- Years shipping ML
I work on the unglamorous half of applied AI: the retrieval layer that actually returns the right chunk, the eval harness that catches a regression before a customer does, and the serving path that keeps p95 latency under budget.
Most recently I have been building agentic retrieval systems — planners that decompose a question, route across heterogeneous sources, and cite what they used. Before that, production computer vision on edge hardware.
I care about measurement. If a change cannot be shown to help on a held-out set, it is a hunch, not an improvement.
Languages & core
- Python96
- TypeScript82
- SQL88
- Rust58
Modelling
- PyTorch93
- TensorFlow / Keras78
- Hugging Face Transformers92
- LoRA / QLoRA fine-tuning88
LLM systems
- LangGraph / LangChain90
- RAG architecture94
- pgvector / Qdrant87
- Evaluation & guardrails89
Platform
- Docker & Kubernetes80
- AWS / GCP82
- MLflow / Weights & Biases85
- CI/CD for models83
Also shipped
More from the lab.
Latest from the blog
Notes from the build.
Long-form writing on LLM architecture, retrieval, evaluation and the unglamorous parts of shipping AI.
2 min read
RAG is a retrieval problem, not a prompting problem
Two-thirds of the accuracy gain in our production RAG system came from changes upstream of the model. Here is the measured breakdown, and what to fix in what order.
1 min read
How to evaluate an LLM system without fooling yourself
Vibes-based evaluation is how teams ship regressions. A practical framework: golden sets, calibrated judges, regression gates, and the failure modes of each.
1 min read
Fine-tuning vs RAG: a decision procedure, not a debate
The question is not which is better. It is which failure you are trying to fix — missing knowledge, wrong behaviour, or unacceptable cost.
Next
Have a model that needs to reach production?
I take AI systems from notebook to a service your team can actually run.
Start a conversation