July 22, 2026
Part 1 built the prototype — vectors, cosine similarity, a pre-trained embedding model. Part 2 builds the real thing: enriched chunks, a pgvector index with HNSW, BM25, RRF fusion, CLIP image embeddings, and a 3-way hybrid.
July 14, 2026
Keyword search fails when meaning doesn't live in the words. This post builds semantic search from scratch — vectors, cosine similarity, and a real embedding model — on Vuori's full product catalog.
April 23, 2026
I typed a request into Claude Code. What actually happened? A deep dive through the inference serving stack and the bottleneck rotation up the supply chain — to the $380M machines that gate it all.
March 20, 2025
How I built a 6-layer AI pipeline that compresses CRM data, call transcripts, and emails into prioritized, actionable intelligence — so a sales rep can go from notification to drafted email in 90 seconds.
February 25, 2025
How I built a 4-stage LLM pipeline that triangulates emails, call transcripts, and usage metrics to produce forensic-quality churn reports — with no agents.