Things I've built.
A runtime and spec for executable markdown skills consumed by AI agents.
One spec for executable markdown skills, one reference runtime in Rust, two content modes (prose-only Soapstones-compatible + typed structured with frontmatter), and an eventual hosted registry where any agent can pull a skill written by any other agent. v0.1 ships parser, validator, local install/list/info; v0.2 lands the hosted registry.
Open-source AI inference logging infrastructure for EU AI Act Article 12 compliance.
A drop-in Cloudflare Workers proxy that sits between an AI app and its provider (OpenAI, Anthropic, Gemini). Every inference becomes an immutable, hash-chained audit record. Raw prompts and outputs are never stored — only SHA-256 fingerprints — so it's GDPR-clean by design. Compliance report exports to PDF in one click.
Designing a controller to steer a simulated car.
Open challenge from Comma.ai: design a steering controller for a simulated car. Engineering exercise in classical control + responsiveness vs stability tradeoffs.
Classifier predicting whether a court case will be dismissed.
Court case outcome prediction from filing text. ML / NLP exploration in the legal-tech space.
Scores AI responses (and prompts) across five dimensions of intelligence.
Scores AI responses on creativity, empathy, problem-solving, worldview, and adaptability. Early exploration in evaluating LLM output beyond raw correctness — adjacent to current AILedger / skillrt thinking on agent behavior measurement.
Tools for calculating a stock's Net Current Asset Value.
Quantitative screener for Benjamin Graham-style net-net stocks (companies trading below liquidation value). Pull, compute, filter.
Auto-buy GME any time Roaring Kitty posts on Twitter.
Half-joke, half-real: scrape Roaring Kitty's twitter, parse for $GME signals, fire a market order. Built during the meme-stock revival as a quick exercise in event-driven trading scaffolding.