PRISM: AI recruitment and proctored assessments
One platform for the whole hiring pipeline: job post, CV screening, proctored online test, review and outcome emails. Built because we were losing hours reading CVs by hand and supervising tests in person.
- Claude screens each CV against the job's requirements and returns a requirement-by-requirement breakdown with evidence, not just a score. It's optional per job, and HR can always screen people in or out manually.
- Live proctoring: candidates wait in a virtual waiting room until HR starts their test, and HR can watch, listen in or talk to them privately over self-hosted LiveKit.
- Cheating detection runs in the candidate's browser with MediaPipe, flagging looking away, a missing face or a second person, plus tab switches, copy-paste and screenshot attempts. No video is sent anywhere for analysis.
- Every violation is stamped with its moment in the recording, so a reviewer clicks it and jumps straight there.
How I built it
Stack. Django REST Framework backend and a React + TypeScript frontend (TanStack Router, Tailwind), self-hosted on an Ubuntu VPS behind Nginx and Gunicorn.
Built for real conditions.
- A TURN relay so video works on restrictive Nigerian mobile networks, with a snapshot fallback when video can't get through.
- Sessions recorded in 30-second chunks, joined with FFmpeg and streamed back through signed, time-limited links.
- The timer runs on the server, so it can't be tampered with. Connections recover from drops automatically, and HR can add time or reopen a test for someone whose internet failed.
- Test links expire for newcomers without cutting off anyone already mid-test.
What I learned. The hard part isn't the features. It's making them hold up on bad networks, cheap laptops and real candidates. Most of the engineering went into that.
