Getting a Small Model to Guess Who Should Meet Whom

NOGnet already matches people at conferences. Attendees fill out a profile: a role, a line about what they're working on, a couple of interest tags. The platform suggests a short list of other attendees worth finding at the coffee table. The logic underneath is about as fancy as you'd guess. Shared tags score points, shared region or employer type adds a bit more, sort, done. It works well enough that no organizer has ever complained about it. I'm not sure that means it's good. It might just mean nobody checks.

Reading profiles instead of scoring tags

For a few weeks now I've been testing something on the side. A small language model running on a machine at home, no cloud calls, just a box in a closet doing inference when I ask it to. The idea is narrow: instead of scoring tag overlap, hand the model two attendee profiles as plain text and ask it to reason about whether these two people might have something worth talking about, and why. Tags flatten a person into a handful of checkboxes. A profile written in someone's own words carries more than that, even a short one, and I wanted to see whether a model could actually use it.

Picture two profiles. One person writes that they run the network for a small regional ISP and are tired of debugging the same peering issue every quarter. Another writes that they just moved into a role building automation tooling for a much bigger operator, trying to convince their team that manual peering configs don't scale. A tag match might connect them on "peering" and call it done. A model reading both might notice the size gap, the seniority gap, the fact that one of them has the exact problem the other one solves for a living, and say something closer to why they'd actually want ten minutes together. That's the hope, anyway. I don't have real numbers yet, and I'm wary of anyone, including myself, who mistakes a good demo for proof.

The parts I'm unsure about

The first honest question is whether any of this beats simpler methods I haven't even tried yet. Embedding the free-text fields and ranking by similarity would be cheaper and would probably catch a good chunk of what a full reasoning pass catches. I keep asking whether the model's explanation does real work or just dresses up a similarity score as a reason. I don't know yet. Testing that properly means building the boring baseline first, which I keep putting off because reading model output is more fun than writing evaluation code.

The second question is harder, and I don't expect to solve it soon. What does a good match even look like? NOGnet has no signal for "these two people talked and it mattered." At best there's a feedback field on a post-event survey that maybe a third of attendees fill out, and even then, "we had a nice chat" doesn't tell me whether the introduction caused anything or whether they'd have found each other anyway in a room of forty people. Conferences are small enough that good conversations happen regardless of what any matching feature does. Proving this adds something on top of that baseline might be close to impossible with the data I have.

There's a third thing that bothers me more than accuracy does. A tag list only knows what someone explicitly checked. A model reading a paragraph infers things nobody said outright: seniority from word choice, frustration from a throwaway line. Some of that inference is exactly the value I'm after. Some of it feels like the model quietly psychoanalyzing someone who just wanted to list a job title to get a badge. I haven't decided where the line sits, and I'm not comfortable putting this in front of real attendees until I have a better answer than "it seemed fine in testing."

Right now this lives on my own machine, against old anonymized profile data, nowhere near production. I read the output by hand, wince at a good chunk of it, and go back to adjusting the prompt. No timeline, no promise this ships. It might turn out a plain similarity score does ninety percent of the job and the model was never the point.

If it goes anywhere, I'll write about what worked instead of what I hoped would work. For now it's a model in a closet, guessing about strangers, while I second-guess the guesses.