Full Stack Engineer

The system

Prospeer was building an AI hiring platform and I headed software development for it. The product is one long pipeline: a posting drafted by a model, outreach to matched candidates, resume scoring, match ranking, an automated interview, and a shortlist at the end.

I owned the chain end to end and built two of the harder pieces myself — a one-way video interview interface on WebRTC, and AI-proctored assessments with anti-UFM checks. Roughly 8,000 candidates were scored in the first quarter after launch, against a manual process that had managed a few hundred.

What was hard

Scoring a recruiter would defend. Anything that ranks people has to be explainable to the person acting on it. We kept the score but always showed the evidence under it: which requirement matched, and where. Recruiters stopped overriding it once they could see why.

One-way video without a media server. Live interviews need infrastructure. One-way recording does not, if you record locally and upload in chunks. That decision removed a whole category of running cost — and the chunked upload and resume logic still took longer to get right than the interview UI itself.

Proctoring without accusing anyone. Every anti-UFM signal is probabilistic. A tab switch is not cheating. Signals went to a human for review and never auto-rejected, which is slower and the only version I would sign off on.

What I would do differently

We built the automated interview before checking that candidates would finish one. Completion sat at 60% for the first month. Watching ten real sessions first would have found what a rewrite eventually did.