Small builds, experiments, and working methods — the things I make to understand the systems I design for. Some live, some in progress, all honest about which.
A React terminal for browsing live prediction markets, with an LLM analyst that produces structured quantitative read-outs of any market: odds, volume, liquidity, time horizon. Built to explore how model-generated analysis should present uncertainty to someone about to act on it.
What it tests How should an interface show a model’s probabilistic analysis without overstating its confidence, when someone might bet money on it?
A Python pipeline using a locally-hosted LLM to analyse hundreds of sensitive support conversations, entirely off cloud infrastructure. The guard: the model had to cite a session ID for every finding, and I cross-checked each citation against the source transcripts by hand.
What it tests Can LLM-assisted research analysis be made hallucination-resistant enough to base design decisions on? Citation-forcing plus manual verification was the answer.
The portfolio is its own experiment: cross-document view transitions, motion that degrades honestly under prefers-reduced-motion, and a 404 page that refuses to hallucinate. Every microinteraction has to earn its place by demonstrating the practice it decorates.
What it tests Can a portfolio about honest AI be enjoyable without overselling anything? This site is the test.
Nothing here is a product pitch. If an experiment fails, it stays — the point is what it taught, not how it looks on a slide.