I built an open-source tool for technical interviews where candidates work with their actual AI coding assistant and the hiring manager sees the full thinking process.The problem: take-home cheating rates doubled to 35% in a year. Live coding has become a surveillance exercise. The industry is stuck in an arms race between stop candidates from using AI and use AI without getting caught. Meanwhile, every engineer actually uses AI at work.InterviewSignal takes a different approach — don t ban AI, instrument it. The tool hooks into the candidate s AI assistant (Claude Code, Codex, Cursor, Gemini CLI, Aider) and captures every prompt, tool call, edit, and iteration. Events are hash-chained so the sequence is tamper-evident. On submit, the hiring manager gets a full transcript of the working session, auto-graded against their rubric.The grading explicitly differentiates high-leverage AI use (candidate decomposes the problem, directs the AI, verifies output) from low-leverage use (throws the problem at the AI, accepts the answer, says yes ). When every candidate produces working code, how they got there is the only remaining signal.Technical details people here might care about: zero external dependencies for core (stdlib only — urllib, hashlib, http.server). Hash chain integrity on every event. Per-prompt git commits for independent cross-verification. Client-is-untrusted security model — detection not prevention, absence of signal is signal. Tamper flags surface anomalies automatically (hooks gaps, diff mismatches, prompt ratios).HM experience is 100% browser (dashboard). Candidates work in their own editor — VS Code, terminal, whatever they normally use. pip install, share a code, done.GitHub: https://github.com/NikhilSKashyap/interviewsignalWrote more about the philosophy behind it: https://substack.com/home/post/p-198234845