Dating matches from a model that shows its work
Dating apps tell you who to date without saying why or how sure they are. I wanted to see what that loop looks like when the matching is grounded in data: you answer a few questions a day, a few matches arrive each morning, and every match says how sure the model is and why it picked that person.
Questions aren't random. For each person I pick the ones expected to shrink the uncertainty in their match scores the most, weighted by how much they care about each dimension. Every answer is a Kalman-style update to a Gaussian belief, which keeps it cheap and keeps every number explainable to the person it describes.
If everyone just sees their own top three, popular profiles end up in dozens of lists and a lot of people get nothing. Daily matches are assigned with greedy b-matching instead: a match is shown to both people and nobody gets more than three a day. It costs a little like rate, and I think that trade is worth it.
There were no users, so I built a simulator first: people with hidden traits, answers with noise, and a random chemistry term no questionnaire could ever capture. The engine never sees the ground truth, and the simulator measures how close it gets. It also showed me where learning from likes helps and where it doesn't.
The belief update runs in TypeScript in the API and in Python in the nightly job. The Python side exports the question bank and recorded answer sequences with their expected results, and the API tests replay them, so the math can't quietly drift between the two.
The simulator writes its numbers to a results file every time it runs, and the charts in the README are drawn from that file rather than typed in. Two answers arriving at once from two tabs are handled under a row lock on the person's belief, and there's a test for that. Scoring city by city was checked against scoring everyone at once and gives exactly the same matches.
Use chat activity as a signal too, not only likes and passes, and replace the hand-written question bank with a proper item response model once there is real data to fit it to.