Knot AI

Knot.ai — a spoken AI interview panel that adapts to you

The problem
Interview practice today is broken. People rehearse by reading questions off a screen and typing or thinking through answers. But real interviews are nothing like that — they're spoken, they get harder when you're doing well, they stop and press you when you're vague, and they're run by several people who each want something different and compare notes afterward. Existing tools can't reproduce any of that, and the ones that give "AI feedback" often just make things up.

What Knot.ai does
Knot.ai runs a full, real-time voice mock interview with a panel of AI interviewers. You pick who you face — technical, hiring manager, product, customer, behavioural — and the order. Each interviewer has its own agenda and voice, and a summary of every round is passed to the next one, so nobody repeats what's already been covered. The interview is genuinely conversational: you can talk over it and it stops, like a person would, it climbs in difficulty as your answers land, and it refuses to move on from a hand-wave until you give something specific.

Our goal: feedback you can actually trust
The whole project is built around one principle — an AI shouldn't be able to make things up about you. So the final assessment follows a hard rule: every claim has to quote you. After the interview, each strength and weakness the model writes is verified against your real transcript, word for word. Anything it can't tie to something you actually said is dropped — and the report shows you how many were cut. No invented praise, no invented criticism.

To make the interview behave the same way every time, the control logic — difficulty, interviewer hand-offs, the "pin down vague answers" rule, and the evidence check — lives in a deterministic server layer, not just a prompt. The AI supplies judgement and phrasing; our server decides what actually happens.

What it demonstrates
Real-time interruptible voice · multiple interviewer roles and personalities · shared candidate context between roles · dynamic follow-up questions · controlled interviewer turn-taking · role-play/scenario questions · adaptive difficulty · detection of vague or contradictory answers · evidence-based feedback linked to the transcript · a structured final assessment · and a clear disclosure that you're talking to an AI.

Built with
Agora Conversational AI (real-time voice), Claude (the interviewer brain), a Fastify + TypeScript control proxy that owns every turn, a React/Vite front end, and SQLite as the transcript source of truth.

In one line: a spoken interview that pushes you like a real panel — and an assessment that can only tell you what you actually said.

 Demo video: https://youtu.be/tPoFIWdfmPA
 Live demo (no sign-up): https://knot-ai.onrender.com
 Code: https://github.com/SpaceWalkerr/Knot.ai
 PPT: https://docs.google.com/presentation/d/1aMVfnVpalEFHetsi8DNOXY1f6QvObXWX/edit?usp=drive_link