ODF Connect
AI cofounder matching for On Deck Founders using hybrid search and LLM-scored compatibility.
The Problem
Founders struggle to find cofounders aligned on values, technical depth, ego, and availability — attributes keyword matching cannot detect. The ODF community needed matching based on nuanced signals beyond surface skills.
What I Built
A two-service platform: Next.js 15 frontend (Vercel) plus Python FastAPI backend (Railway). The flow: a stateful chat-based interview (Claude Sonnet) extracts a structured persona; the system synthesizes founder profiles into signals, embeds them into dense vectors, retrieves candidates via hybrid KNN + BM25, scores them against the user's persona with Claude Opus, and generates tiered match narratives with concerns and overlap topics. Backend uses SQLite + sqlite-vec for vectors and FTS5 for keyword search. Three tiers of recommendations with editable personas and a ranked match feed.
Notable
A "Mutual Fit" signal extracts what each founder is seeking (from synthesis), not just what they are — enabling bidirectional compatibility scoring. Profiles synthesized once with GPT-5.2 (cached by content hash), embeddings cached with OpenAI text-embedding-3-small, then scoring runs in parallel batches with Claude Opus to handle rate limits while keeping costs predictable (~$0.18 per match query).
Stack
Status
Live at odf-connect.vercel.app. Open source at github.com/blaizew/odf-connect.