I'm a data engineer at Autodesk in San Francisco. Here are a few side projects, most interesting first.
| Project | What it is | Measured |
|---|---|---|
| Nightshift | A live benchmark: can an AI agent make money trading prediction markets, and get better at it by rewriting its own code? Claude agents trade real Polymarket and Kalshi order books for 28 days; half keep their own code between sessions, half are reset. | v0.1 live Oct 27 – Nov 24, 2026. 6 accounts, real fees, every number with a 95% interval. The pilot fund it replaced went +2%, then −57% in ten days. |
| ShingleBench | A benchmark for whether an AI model can start a real business: it builds the site, brand, ads and pricing in a sandbox, runs it for simulated weeks, then LLM judges and blind human graders score it. | 26 briefs in 6 categories. The demo round separates an honest scripted founder (6.18/7) from a corner-cutting one (2.25/7), with non-overlapping 95% CIs. |
| HealthForce | An ER assistant on Snowflake and Claude: identifies an unconscious patient from what the team can see and opens their chart from every care site. | 🏆 Won the Intelligence at Work Hackathon (Snowflake × Anthropic). 1.42M patients to one likely match with a single bedside check. |
| vendormatch | Entity resolution for messy vendor records. Cheap matchers take the easy pairs; a local LLM only sees the hard ones. | F1 0.936 vs 0.906 for the best string match. A local Qwen3-14B on 3.6% of pairs cuts false matches 37%. 50k records in 31 s. |
| NanoServe | A readable LLM serving stack for Apple Silicon: OpenAI-compatible API, streaming, benchmarks. | Continuous batching: 2.2x throughput, 4.5x lower worst-case TTFT for 8 Qwen3-14B streams. About 5 ms of server overhead. |
| NanoTrain | Train a small language model from scratch, then serve it with NanoServe. | 4.9M-parameter model to 1.56 nats/byte on Tiny Shakespeare in 11 minutes |
| Terra | A robot arm that walks: an SO-101 arm on a 3D-printed quadruped, designed and stress-tested in Autodesk Fusion. | Arm mount redesigned from a 1.0 to a 3.1 safety factor under a sideways load (Fusion simulation) |