AI Engineer / San Francisco
I build AI agents, agentic workflows, and context-engineered systems that learn from every human interaction and get smarter in production.
Powered by my actual code, projects, and work historyFounding AI Engineer / Oct 2024 - Present / San Francisco
Sole AI engineer. Took the entire AI platform from 0 to 1, now scaling 1 to 10. Built an intelligent bookkeeping system that classifies transactions across hundreds of account categories at 94% production accuracy, up from 58% before my system. Separate modules handle counterparty categorization and transfer matching. The platform learns from every accountant correction, auto-approves high-confidence predictions, and routes edge cases to human review. End-to-end pipeline from bank feed ingestion through agentic classification to automated QuickBooks sync.
Built an autonomous meeting bot that joins, records, and transcribes across Teams, Meet, and Zoom simultaneously. 10+ concurrent sessions, 46% compute reduction.
Predictive maintenance on sensor data. Reduced downtime by 28% and extended vehicle life by 19%. Built OpenCV path-finding for autonomous agricultural navigation.
Full-stack on production client products. React/Redux frontends, Node.js/GraphQL backends, DynamoDB storage layer.
Cancer detection models under 100MB with clinical-grade precision. Warehouse CV tracking on GCP, shipped in under a week.
Bone age prediction from radiographs using depth-wise separable convolutions. Published at IEEE ICSES 2021.

Your 3rd move deletes your 1st. The game never draws. Real-time multiplayer over Socket.IO with private rooms and Bo3/Bo5 series.

Detects slouching via MediaPipe and fires a water gun at you. Servo tracks your face for aiming. Arduino serial control.

Chatbot that remembers across sessions. No chat history sent to the LLM. All continuity comes from Mem0's semantic memory layer.

Deep Q-Network learns Snake from scratch. Custom Pygame engine, 11-state representation, experience replay. Zero human strategy.

Webcam reads your expression, CNN classifies mood, K-Means clusters Spotify tracks by audio features. Serves matching songs.

Knowledge graph over 610K MoMA + Met records. SPARQL queries on RDF triples in Stardog Cloud. No ORM, pure graph queries.

YOLOv8 Large exported to ONNX with hand-rolled NMS in pure NumPy. FastAPI serving, 0.885 mAP@50, fully Dockerized.

The site you're on. RAG-powered AI chatbot that retrieves from 400+ embedded docs of my actual work to answer questions about me.

Benchmarking 8 embedding models (MiniLM, MPNet, E5, Instructor-XL) on cosine similarity search. Systematic model selection study.
Drag to explore
AI engineer in San Francisco. I build production AI systems: multi-agent architectures, agentic workflows, context engineering, and self-improving pipelines that learn from human feedback at every step.
At Legix, I designed the entire AI platform as the sole engineer, from classification engines to evaluation frameworks to observability stacks. Before that, I shipped CV models for cancer detection, built autonomous meeting bots, trained RL agents from scratch, and wrote path-finding for autonomous vehicles. I care about AI that works in the real world.
Always interested in challenging problems and new opportunities.