DealCollab

B2B deal discovery was almost entirely driven by manual networking, cold email, and conference attendance — high-cost activities with unpredictable conversion. Buyers and sellers in niche verticals had no way to describe what they were looking for beyond keyword tags, which meant a distributor seeking "sustainable FMCG suppliers in Southeast Asia" would receive the same results as one searching for "bulk commodity trading" if both checked the same category box.

The platform's founding team had validated the problem with 40+ business owners: 78% said they'd found their best partnerships through warm introductions, and 91% said cold outreach from unknown parties almost never converted. The gap was a semantic understanding layer — something that could read a company's profile and infer partnership fit the way a well-connected industry broker would.

The matching engine is built on pgvector (PostgreSQL vector extension on Supabase). Company profiles, partnership briefs, and deal requirements are embedded using OpenAI's text-embedding-3-large model into 1,536-dimension vectors. When a user initiates a match search, their profile vector is compared against the corpus using cosine similarity with HNSW indexing for sub-50ms retrieval at scale, then re-ranked by a Node.js scoring service that applies weighted business rules (geography, deal size, exclusivity requirements).

The backend is a Node.js/Express API with a well-defined service layer. Each match goes through three passes: vector similarity (retrieval), rule-based filtering (hard constraints), and an LLM re-ranking step where Claude evaluates the top 20 candidates and returns a compatibility summary with reasoning — this is what gets shown to the user as the "why we matched you" explanation card.

The frontend is a Next.js application with Supabase Realtime powering the collaboration workspaces. Once two parties mutually accept a match, they're dropped into a shared workspace with document exchange, milestone tracking, and a conversation thread — keeping the deal workflow inside the platform rather than scattering it across email.

The platform facilitated 200+ accepted deal connections in its first operating quarter, with an average match-to-response rate of 64% — more than 3× the cold outreach benchmark the founding team used as their baseline. Of those connections, 38 progressed to signed agreements within the first six months.

Semantic matching significantly outperformed keyword-only search in blind tests run during QA: users rated AI-generated matches as "highly relevant" 71% of the time versus 29% for keyword-filtered results on the same dataset. The LLM-generated compatibility summaries were cited by users as the single most useful feature — they reduced the time spent evaluating a potential partner from ~15 minutes of independent research to under 2 minutes.