GrilledAI

Practice interviews with an AI interviewer, then get honest, specific, rubric-scored feedback — the opposite of "Great answer, 9/10!". Built feedback-first (see docs/).

Stack

• Next.js 16 (App Router) + TypeScript + Tailwind
• Google Gemini (free tier) via @google/genai — provider-agnostic wrapper in lib/llm.ts, so swapping to Claude later is a one-file change.

How it works

1. Setup (/) — enter the role, interview type, job description, and resume.
2. Interview (/interview) — a live text chat. The interviewer (one Gemini prompt) asks one grounded question at a time and probes weak answers.
3. Feedback — on "End", a separate, cold evaluator prompt scores every answer on 6 dimensions, names the biggest weakness, and shows a stronger rewrite.

Two distinct LLM roles live in lib/prompts.ts — the encouraging interviewer and the honest evaluator are deliberately separate. That separation is the moat.

Run it locally

1. Get a free key at https://aistudio.google.com/apikey
2. Put it in .env.local:

   GEMINI_API_KEY=your_key_here
   

3. Install + run:

   npm install
   npm run dev
   

4. Open http://localhost:3000

Project docs

• docs/evaluator-design.md — the rubric + evaluator prompt (the core asset)
• docs/sample-transcripts-swe.md — test fixtures (weak/mixed/strong)
• docs/proof-runbook.md — the $0 manual validation of the concept

Roadmap

• Phase 0 — prove the evaluator beats naive feedback (done, $0)
• Phase 1 — web app: setup → text interview → report (this)
• Phase 2 — voice input (browser Web Speech API) + delivery stats
• Phase 3 — auth, DB, history, role templates
• Phase 4 — photorealistic avatar (Tavus), swapped in at the front