Matchmaking, done right: AI in B2B networking

Why the old model is broken

The filter-and-list approach treats matching as a database query. It is not. The variables that determine whether two people should meet at an event — current project, recent purchase, hiring intent, last quarter’s miss — are not in any structured field. They’re in the email signature, the LinkedIn post, the conference badge from last year.

The right meeting at a trade show is rarely the obvious meeting. The obvious meeting is on the agenda. The right meeting is what changes the quarter.

What changed

Two things, technically. First, embeddings on attendee profiles let us reason about similarity of intent, not just identity match. Second, retrieval-augmented generation lets the matchmaker explain why a meeting is recommended — which is often more important than the recommendation itself.

{% include post-callout.html variant=”gradient” title=”The pitch in one line” body=”A good matchmaker doesn’t tell you who to meet. It tells you why this meeting matters before you walk into it.” %}

What to actually build

1. Profile enrichment from a single LinkedIn URL — title, company, recent posts, current focus.
2. Pairwise relevance scoring across the attendee pool — embeddings, not filters.
3. Per-meeting context briefs — three lines explaining why the model thinks this pairing is high-value.
4. Calendar integration that respects the attendee’s actual day — the model proposes, the attendee disposes.

Skip any of the four and you have a list, not a matchmaker.