~/lab/experiments
Two kinds of entries live here: deployed experiments you can poke at right now, and active research shown with its real status. Nothing on this page claims to be more done than it is.
A retrieval pipeline over everything on this site. Ask it something — it answers only from what's indexed, and says so when it doesn't know.
hypothesis: A portfolio that can answer questions about its owner is a better demo of retrieval engineering than a paragraph claiming retrieval engineering.
Build-time chunked index over resume, case studies and posts; cosine or BM25 retrieval; Llama 3.1 8B on Workers AI; IP rate limiting. It is the chat box above.
hypothesis: An ensemble of Poisson, XGBoost, LightGBM and Elo models, run through 10,000-tournament Monte Carlo simulation, can produce calibrated championship probabilities.
Squad ingestion pipeline and simulation dashboard are running; validation against real tournament outcomes is the open question — which is why this is a lab entry, not a mission.
hypothesis: Tool-using LLM agents can automate the boring half of football-data curation (fetching, reconciling, flagging anomalies) with a human approving diffs.
Next in the queue once the prediction engine ships its first validated cycle.