Our matching is not machine learning, on purpose

Every few months someone asks why MyRoomie's compatibility engine is not a trained model. It would be an easy thing to say. It would demo well.

It is a rule matrix, and that is a decision rather than a stage we have not reached yet.

What the decision is actually about #

The matcher weighs somewhere north of fifty lifestyle and behavioural factors and produces a score. A learned model could produce a score too, probably a better one on some metric I could pick in advance.

The difference is what happens when someone asks why.

With a rule matrix I can answer. You scored low against this person because your sleep schedules are six hours apart, because one of you has a cat and the other declared an allergy, because one of you wants guests most weekends. Every term is inspectable and every term is a thing a human being recognises about their own life.

With a trained recommender I can show you a number and a shrug.

Why that matters more here than elsewhere #

If a music recommender is wrong, you skip the track.

If a housing recommender is wrong, someone signs a twelve-month lease with a person they cannot live with, in a market where moving again is expensive and slow. The cost of a bad match is not symmetrical with the cost of a good one, and that asymmetry should show up in the engineering.

There is a second reason, less philosophical. A rule matrix can be wrong in a way you can fix this afternoon. If we learn that noise tolerance matters more than we weighted it, that is a number in a table. Retraining is a different kind of afternoon.

Where the models do live #

This is not a position about machine learning. The property data infrastructure — FairRent, inside PropertyOS — is genuinely model-heavy, because pricing across twenty-eight markets is a problem where you want the model to find structure you did not know to look for, and where being wrong means a slightly off estimate rather than a bad year.

Different problem, different tool. The interesting engineering decision is almost never can we use the model — it is what does being wrong cost here.

The honest caveat #

A rule matrix has a ceiling. It only knows what we thought to ask, and people are worse at self-reporting than they believe. We collect outcome data and we will learn from it.

But whatever we learn will go back into terms someone can read. The explanation is not a feature we bolted on. It is the product.