In Loop Engineering I pointed an AI agent at one number for my side project, FindHYROXPartner: confirmed matches — two athletes who both say yes to racing HYROX doubles together.
I assumed the hard part was fit. Pair the 5:00/km runner with the 5:00/km runner, tune the scoring, done.
In September we made 76 introductions and got 2 confirmed partnerships. Under 3%. Fit wasn't the problem.
1. Ask "why" at the moment of "no"
Every decline now triggers a one-tap question: why wasn't this a fit? 19 of the first 37 answers said "wrong race or location" — and 16 of those involved someone with no race picked, often paired with a partner in another country or on another continent. A flexible athlete in Ireland got a partner racing in Tampa.
The scoring model was fine. Our location data was broken: half the home cities didn't resolve to a country, so "unknown" was treated as "close enough." Fixing the data did more than any weight tuning ever could.
2. Silence is a "no" people are too polite to say
Of the intros that quietly expired, 41% were one-sided: one person said yes and then waited nine days for nothing. 15 people opened their intro three or more times and never tapped either button.
That's not low engagement. That's a no that costs too much to say. So we made declining kind ("Not for me — they're only told it didn't work out"), put a real deadline in every reminder, and released one-sided matches after ~6 days so the person who said yes gets rematched fast.
3. In a lopsided marketplace, ask the scarce side
Mixed doubles had 39 men waiting for 8 women. Instead of buying more men, we asked 18 women already in our pool one question: would you also race mixed? 3 said yes within nine hours — one was confirmed with a partner four days later.
The cheapest supply is the supply you already have, one question away.
4. Boring bugs beat clever features
55 athletes were silently stored against races that had already happened, and 12 of them were auto-paused. 26 people trying to sign up for a second race were blocked by "you've already signed up." The calendar fix alone grew the active pool from 93 to 105 overnight. No model, no experiment — just reading the data end to end.
What moved
Confirmed partnerships went from 0–1 a week in September to 3, 5 and 6 in the first three full weeks of October: 16 confirmed out of 141 intros, roughly 4× the September rate.
Not everything worked. A redesigned signup screen looked like a 15-point win after a week. Ten days in, it's 61% vs 58% — too close to call. It's still running.
The lesson I keep relearning: before you optimize who you match, make sure people can actually show up together — and that saying no costs them nothing.