i built a sourcing agent that has context on 60,000 past applicants from my ATS

im the talent lead at a fintech, 50 roles to fill in six months. I trained a model that sources based on 60,000 past applicants in my ATS with their interview and hiring outcomes attached. It basically takes an enriched profile and predicts whether that person passes the interview and gets hired. the loop per role goes like this. the agent writes itself a list of hypotheses about which queries will surface people the model scores 8 or above. it runs one hypothesis, pulls 25 results, enriches them, scores them, and tracks what each 8 cost it. if that hypothesis is paying it keeps digging, if its not it drops it and moves to the next one. What was surprising to me was that after about 300 credits it goes back over all the 8s and hunts for patterns in them. For one role, it worked out that alumni of one particular bank in Brazil were scoring well and started targeting them on its own. i didnt tell it do that btw, mind blowing stuff for a non technical person tbh. tech stack: Claude Code - runs the loop on a schedule, writes and scores the hypotheses. Crustdata - people search and enrichment, the data points the scoring model reads. api and cli available. Greenhouse - export of past applicants and their outcomes, the training data and the feedback loop. Scoring model - trained on our own 60k, predicts interview pass and hire from an enriched profile. Over the last month, I've found 30 incredibly qualified candidates, 8 for which offers have gone out and plenty more in the pipeline. caveats, because this is not clean. the model only knows what our past hires look like, so its biased toward exactly that. some roles have so little history in the 60k that there is nothing to train on and we are back to doing those by hand and giving extensive feedback to the agent to train it.