Crucible 🎯

AI-Powered Technical Interview Agent

A conversational AI agent that conducts personalized, adaptive technical interviews based on a candidate's actual learning journey through a 31-day AI engineering cohort — then delivers structured, actionable feedback.

🚀 Live Demo  |  📄 AI Usage Log

---

📌 The Problem

After completing an intensive AI engineering cohort, learners can build the systems — but struggle to confidently explain the engineering decisions behind them in a real interview setting. Crucible closes that gap: it's an interviewer that already knows exactly what each candidate built, what they struggled with, and what they skipped — and asks accordingly.

---

🚀 Live Demo

https://crucible-ib0j.onrender.com/

⚠️ First request may take up to a minute. This is deployed on Render's free tier, which spins the server down after periods of inactivity. The first request wakes it back up; every request after that is fast.

---

✨ What It Does

• 🎯 Personalizes to each candidate — pulls their actual completed missions, attempts, and skipped topics from candidates.json to decide what to ask
• 💬 Conducts a real multi-turn interview — minimum 8 questions, spanning at least 4 distinct curriculum days
• 🧠 Generates genuine follow-ups — every next question reacts to what the candidate just said, not a fixed script
• 🧵 Maintains full conversation context — session-based, so the model always knows what's already been asked and answered
• 📋 Produces structured feedback on completion — a summary, strengths, gaps, and next steps, generated from the actual conversation

---

🛠️ Tech Stack

Category	Tool
🐍 Backend	Python 3.11, FastAPI
⚡ LLM Inference	Groq (Llama 3.3 70B, OpenAI-compatible SDK)
🌐 Frontend	Vanilla HTML/CSS/JS (no build step)
☁️ Deployment	Render
🗃️ State	In-memory session store (per technical spec — no persistence required)

---

📁 Project Structure

crucible/
├── app/
│   ├── main.py            # FastAPI app, /api/interview route, root redirect
│   ├── llm.py             # Groq client — JSON-mode enforced, retry + fallback
│   ├── prompts.py         # Builds the interviewer system prompt from curriculum + candidate data
│   ├── session_store.py   # In-memory per-session conversation state
│   └── models.py          # Pydantic request/response schemas
│
├── data/
│   ├── curriculum.json    # 31-day cohort curriculum
│   └── candidates.json    # Candidate mission history and learning signals
│
├── static/
│   └── index.html         # Chat UI — candidate picker, live conversation, feedback panel
│
├── PROMPTS.md             # Full AI usage log
├── requirements.txt
└── .env.example

---

🔌 API Contract

Implements the single required endpoint exactly as defined in the technical spec: POST /api/interview Start a session:

{
  "sessionId": "abc-123",
  "candidate": { ...candidate object from candidates.json... }
}

Continue a session:

{
  "sessionId": "abc-123",
  "message": "candidate's answer"
}

Response (mid-interview):

{
  "reply": "...",
  "done": false
}

Response (interview complete):

{
  "reply": "...",
  "done": true,
  "feedback": {
    "summary": "...",
    "strengths": ["..."],
    "gaps": ["..."],
    "next": ["..."]
  }
}

No authentication required, per spec.

---

🧠 How It Works

1. On the first request, the candidate's passed missions are cross-referenced against curriculum.json to build a personalized system prompt — the model only asks about topics the candidate actually completed.
2. The model is instructed to ask one question at a time, generate a genuine follow-up before switching topics, cover at least 8 questions across at least 4 curriculum days, and adapt difficulty based on the candidate's attempt counts and signals.
3. Every response is forced into strict JSON via Groq's response_format: json_object, with retry logic and a graceful fallback if the model ever returns malformed output — so a parsing hiccup never crashes the interview mid-conversation.
4. Once the model determines enough ground has been covered, it sets done: true and generates the structured feedback object in the same turn.

---

▶️ Run Locally

git clone https://github.com/mysticalayushi/crucible.git
cd crucible

python -m venv .venv
.venv\Scripts\Activate.ps1        # Windows PowerShell
# source .venv/bin/activate       # macOS/Linux

pip install -r requirements.txt

cp .env.example .env              # then add your own GROQ_API_KEY
uvicorn app.main:app --reload

Open http://127.0.0.1:8000/ — it redirects to the chat UI.

Get a free Groq API key at console.groq.com — no credit card required.

---

⚠️ Known Limitations

• Cold starts on Render's free tier (see Live Demo note above)
• No persistence — conversation state lives in memory and resets on server restart (intentional, per the spec's "out of scope" list)
• Single-process session store — fine for this scope; a production version would move to Redis or similar for multi-instance deployments

---

🔭 Future Improvements

• Move session state to Redis for horizontal scaling
• Add streaming responses for a more natural typing-indicator feel
• Voice input/output
• Persist interview transcripts for candidate progress tracking over time

---

📋 Project Info

Field	Detail
👩‍💻 Built by	Ayushi Rai
🏆 Event	ABTalks Vibe Code Hackathon
🤖 AI Tool Used	Claude (Anthropic) — full log in PROMPTS.md
📅 Date	August 2026

---

_{Built solo, end-to-end, during the hackathon build window.}