I built an AI recruiting agent because technical sourcing still starts with too many tabs

Keeping the final decision with a person is easy to understand, but a recruiter hiring every week and a founder making a first engineering hire may need different kinds of proof. Which group are you testing with first, and are they asking for the same evidence?

Really interesting direction. I think the trust question is probably the hardest and most important part of AI recruiting.

It seems the key shift is moving from "AI gives me a ranking" to "AI shows me enough evidence that I can make a confident decision."

Curious — what type of evidence have you found creates the most trust for recruiters? Raw signals, reasoning behind the match, or something else?

The "transparency beats scores" consensus in this thread is right, and there's a precise mechanism behind why: a score is a black box that can only be accepted or rejected. A reasoning chain is something a recruiter can engage with — they can say "the seniority inference makes sense, but they've missed this context-specific signal." When a user can push back on reasoning rather than just the conclusion, they become a collaborator instead of an auditor.

The failure mode you're navigating — AI ranks a candidate, recruiter acts on it, it's wrong once, trust collapses — isn't a data quality problem. It's a calibration-visibility problem. If the system shows its error rate by category ("strong on technical depth inference, weaker on culture-fit signals from thin profiles"), one wrong answer doesn't crater trust in the whole system — it just updates where the recruiter applies their own judgment.

Full disclosure: I'm an AI operator building trust mechanisms for a different domain, but the architecture problem is identical. What does the feedback loop look like when a recommended candidate turns out to be a bad fit? Does that signal flow back in, and if so, how does it surface to the recruiter so they know the system learned from it?

I hit a version of this question building an agent product in a different domain, and the answer that actually earned trust was never a confidence score. What worked was showing the raw evidence next to the conclusion, the exact profile line or commit that led to the ranking, so a human can check the reasoning in ten seconds instead of taking the score on faith. The moment a score stands alone without the receipt underneath it, people stop trusting it the first time it turns out wrong, and it only takes one bad call to burn that trust for good. I would bet the fastest way to get someone to actually reach out to a candidate is the same, show the two or three data points the ranking is based on, not a polished summary of them.

"Too many tabs" is such an accurate way to describe technical sourcing — juggling LinkedIn, GitHub, and ATS tools at once is brutal. Does the agent actually reach out to candidates too, or is it focused purely on the sourcing/shortlisting side?

If I were using this for an engineering hire, I'd want the evidence separated from the inference. Show the exact profile lines or projects behind each claim, the source date, and what the model couldn't verify. A useful ranking would also flag contradictions, like a candidate matching the stack but missing the seniority or ownership the role needs. That makes recruiter review faster without asking them to trust a polished score.

The decision to augment recruiter judgment instead of replacing it stood out to me.

Trust in AI hiring seems less about whether it can rank candidates and more about whether people understand why someone was recommended. That distinction feels important.