esci

Self-hosted AI product search for e-commerce — embeddings + reranker trained on 427k real Amazon relevance judgments. Runs on CPU. No per-query fees.

Search evaluation with Jev as the judge. turbopuffer does the retrieval and hosted reranking, and DuckDB keeps every run and computes the metrics in SQL.

Hybrid retrieval pipeline for e-commerce search, scored against Amazon's ESCI human relevance judgments: 8,956 queries, graded nDCG@10, paired significance tests on every stage delta.

A JavaScript implementation of the ESCI software simulator used in Introduction to The New Stastistics (by Geoff Cumming and Robert Calin-Jageman)

Agent Skill that pre-screens journals against the Web of Science ESCI quality criteria and predicts the likely evaluation outcome

Reproducible IR benchmark for Hinglish / code-mixed product search on Amazon ESCI: BM25 vs multilingual dense vs LLM reranking, evaluated with nDCG/MRR + failure analysis.