A full ML/AI suite for receipt OCR: data prep with WebDataset sharding (200 images → 550 GB), a multi-model evaluation harness (Donut, Qwen2-VL, TrOCR, Claude), a Donut finetune, and a FastAPI inference service. Python | A multi-agent Jedi Council deliberation system built on LangGraph — persona agents debate in public and private rounds, orchestrated through a shared graph, with a retro RPG overlay on top. Python | A custom OpenClaw plugin: a conversational, multi-source ML research radar with LLM-judged relevance that learns from your in-chat 👍/👎 feedback over time. TypeScript |
End-to-end Document AI pipeline for handwritten form extraction, with a human-in-the-loop feedback loop that turns corrections into better models. Python | Multi-agent equity-research system with verification gates and citation tracking, built on LangGraph + Postgres + LangSmith and served via FastAPI. Python | Classifies StarCraft II players by skill tier from gameplay telemetry — feature engineering and supervised modelling on behavioural data. Python |
Production-ready FastAPI + ML template: versioned routers, per-version Swagger docs, uv dependency management, and Docker out of the box. Python | An interactive visual git-history explorer with sliders to scrub through commits — built entirely with Claude Code. JavaScript | Open-source Claude Code skills, agents, and AI tooling — reusable building blocks for agent-driven engineering. |