Work
Selected projects and write-ups.
RedInk // multi-agent grading, with a human in the loop
Four LangGraph agents that grade 15+ file types, from PDFs to code to Jupyter notebooks, with a teaching assistant reviewing the calls before anything is final. Full-stack React and FastAPI, deployed on Google Cloud Run. It took grading from around 15 minutes down to 3 per student.
Note: the orchestration was straightforward. The review step was not. Most of the work went into making the handoff between agent and reviewer usable.
Read the write-upSelected work
Status / buildAuto Bench
A benchmark suite for LLM agents built on MCP tools, with scoring rubrics that surface where a model fails.
Cognitive Canvas
A production MCP server pairing GPT-4o-mini with a ChromaDB store, so Claude and Copilot can query a live knowledge base. Pydantic contracts keep the tool calls type-safe.
DejaVu
Near-duplicate video detection with CLIP and ResNet-50 to catch re-uploads of banned content across 884 videos. 0.905 mAP.
AI Adoption and Firm Performance
A panel study on whether AI adoption predicts revenue growth across S&P 500 firms. It does not, beyond ordinary financial controls.
Personal tools
Built for my own useNot everything I make is meant for a resume. If a part of my day is annoying enough, I build a small app to fix it. A running list:
75 Hard tracker
A daily checklist and streaks so I get through the challenge without losing count.
Workout tracker
Logs lifts, sets, and progress so I stop keeping it all in my head.
Meal planner
Plans a week of meals and builds the grocery list that comes with it.
Solving small problems
The list keeps growing. and the sky is the limit.