Projects

Public experiments around language, search, embeddings, and neural representations.

Retrieval

Retrieval ExplorablesNineteen interactive explanations of how search behaves: BM25 and idf, similarity metrics, rank fusion, approximate nearest neighbours, evaluation, reranking, filtered vector search, diversity, ColBERT, hubness, the single-vector limit, and seven interpretability constructions on authored data. Every page states its formulas, works the numbers out on data you can read, and ends with exercises and sources. No dependencies; every number is computed in the browser.

LLMs

Words BetweenAn on-device latent-space instrument. Drag between two words and read what lives in between. Built on Linus Lee’s Contra Bottleneck T5, runs entirely in the browser.Representation ConvergenceOngoing essay series on what survives under transformations of neural representations. Read Part 1 in the blog; lab repository is private until the series is further along.