← All projects
MVP Complete2026 · Co-built (2-person team)

Sift

Drop a widget on your site and turn a pile of technical PDFs into searchable, chat-able, cited knowledge.

Next.js 16Cloud Run workerSupabasepgvectorGeminiRAGDocker

Overview

Industrial companies drown in datasheets, manuals and spec sheets. Sift ingests them through a hardened extraction pipeline and exposes three surfaces: an embeddable widget for the customer's own site, a dashboard, and a full operations portal.

It does real retrieval — streaming, cited answers over your corpus — plus hybrid keyword and semantic search. The economics are engineered: identical PDFs across tenants are de-duplicated by content hash and share one extraction with reference counting, roughly 90% off the AI bill. And it's a business, not a parser — the admin side does customer health-scoring, cost analytics, and email lifecycle automation.

Highlights

  • An extraction pipeline with a circuit breaker, request idempotency, and quality validators.
  • Retrieval on a vector index with an HNSW graph; hybrid keyword and semantic ranking.
  • A containerized worker draining a Postgres job queue, separate from the web app.
← All projects