How Event Matchmaking Algorithms Actually Work

How does simple tag matching work, and why is it weak?

The most common event "matching" is tag matching: attendees pick interests from a checklist, and the system pairs people whose tags overlap. It's cheap and explainable, but weak — it only sees the labels people selected, misses anything not on the list, treats "marketing" as identical whether someone needs it or offers it, and is easy to game. Overlap of tags is a poor proxy for whether two people should actually talk.

Tag matching answers "who checked the same boxes?" when the useful question is "whose needs and offers fit together?" — a question a checklist can't reach.

What does LLM profile enrichment add?

Profile enrichment uses a language model to expand short survey answers into a fuller representation and to extract what each person is seeking versus offering. Instead of a static tag, "I run a bootstrapped SaaS and need help with paid acquisition" becomes structured signal: offers = bootstrapped SaaS operating experience; seeks = paid-acquisition expertise. This directionality — seek versus offer — is what tag systems throw away and what makes complementary matching possible.

How do vector embeddings match meaning?

An embedding turns a piece of text into a vector of numbers positioned so that semantically similar text sits close together in space. Once each person's seek and offer keywords are embedded, matching becomes a nearest-neighbour search: find the offers nearest to this person's needs. Crucially, embeddings match meaning, not words — "needs distribution" lands near "runs a large newsletter" even though they share no vocabulary, a link no tag system would ever draw.

EventIntro stores these embeddings in pgvector and queries them per attendee; the how-it-works page walks the full pipeline.

How is matching different from group formation?

Matching scores pairs; group formation assembles tables. They're distinct problems — two people can be a strong pair while a group of six is unbalanced. Group formation takes the pairwise scores as input and solves for good groups: spreading complementary fits so each table has range rather than clustering all the similar people together. Good breakouts depend on the group-level solve, not just on stacking high-scoring pairs.