Matthew Wang

A shared memory so that when several people each run their own AI agent, those agents can see what the others have already decided.

• Append-only facts, never edited in place, so the audit trail is free.
• Cheap graph and vector filters run first; an LLM only sees pairs that might actually collide.
• Access control lives in Postgres row-level security, not just application code.

TypeScript · Node.js · MCP · PostgreSQL + pgvector · Docker

A five-layer fraud pipeline where an LLM decides which detectors to run, then a person always makes the call.

• A planner chooses which layers to run and how to weight them; routine cases never touch a model.
• Every layer reads the same append-only event log.
• The model recommends, with evidence. It does not freeze an account.

Python · FastAPI · React · Vite · WebSockets · NetworkX

Robot telemetry into diagnostics and a 3D replay, built in a weekend. Best Use of MongoDB at UTRA Hacks 2026.

• Ingestion sniffs the shape of incoming data so a hardware change does not silently drop records.
• Diagnostics are computed on demand, not on the write path.
• The LLM coaching layer can fail; the metrics still come back.

Python · Flask · MongoDB · React · Three.js

A CNN chess engine trained on 5,000+ master games, then improved past that baseline with self-play.

• Supervised pretraining first, then self-play — not AlphaZero from random weights.
• Inference is a FastAPI service on Modal’s GPU cloud.
• No published Elo; the claim is that self-play beat the supervised baseline.

Python · PyTorch · FastAPI · Modal

The arcade cabinet on this site. A maze game rendered live onto a curved CRT.

• The game is a 120 Hz TypeScript loop. The cabinet is just a view.
• The CRT is a shader on a curved mesh, not a post-process.
• Coins to play; three more are hidden in the page.

TypeScript · three.js · Canvas 2D · Astro