README.md

metadata

title: AI Interview Mentor
emoji: 🎯
colorFrom: indigo
colorTo: purple
sdk: docker
pinned: false

AI Interview Mentor

Live at → https://adesh01-ai-interview-mentor.hf.space

An AI-powered mock interview platform for students and instructors. Instructors upload question banks, students take adaptive AI interviews and receive scored feedback.

Demo Credentials

Role	Email	Password
Instructor	adeshboudh16@gmail.com	Adesh@123
Student	adesh@gmail.com	adesh@123

Features

• Instructor portal — create batches, invite students via class code, manage topics, upload question banks via CSV, lock/unlock topics
• Student interviews — AI asks up to 5 questions per session, adapts with follow-up questions on shallow answers, randomised question order each time
• Scored reports — LLM judge evaluates the full Q&A transcript and produces a score (0–100), summary, and improvement tips
• Gemini + Groq fallback — Gemini 2.0 Flash as primary LLM, Groq Llama 3.3 70B as automatic fallback
• Persistent sessions — LangGraph + PostgreSQL checkpointing resumes interrupted interviews

How It Works

1. Instructor signs up, creates a batch, and shares the class code
2. Instructor uploads a CSV question bank and unlocks topics for interviews
3. Student signs up with the class code and picks a topic to interview on
4. AI conducts up to 5 turns — asking, evaluating, and probing deeper on weak answers
5. Interview ends with an inline score and summary; full history saved to the dashboard

Tech Stack

Layer	Technology
Frontend	React 19, TypeScript, Tailwind CSS v3, Zustand
Backend	FastAPI, LangGraph, asyncpg
Database	NeonDB (PostgreSQL + JSONB)
LLM	Gemini 2.0 Flash (primary) / Groq Llama 3.3 70B (fallback)
Auth	Custom JWT (HS256, 15 min access + 7 day refresh)
Deploy	Docker on Hugging Face Spaces (port 7860)

Environment Variables

Set these as secrets in HF Spaces → Settings → Variables and secrets:

Variable	Description
NEON_DB_URL	NeonDB PostgreSQL connection string
JWT_SECRET	Random secret, min 32 chars (python -c "import secrets; print(secrets.token_hex(32))")
GEMINI_API_KEY	Google AI Studio API key
GROQ_API_KEY	Groq API key (fallback LLM)
APP_URL	Deployed Space URL, e.g. https://adesh01-ai-interview-mentor.hf.space

Local Development

make install      # install all backend + frontend deps
make schema       # apply DB schema (run once on fresh DB)
make dev          # start backend on :7860 with auto-reload
make frontend     # start frontend dev server on :5173
make test         # run pytest suite

Docker

make build        # build frontend + Docker image
make docker-run   # run container with .env on localhost:7860

# Reset DB and re-apply schema (wipes all data + checkpoints)
make clean-db

CSV upload — use the Instructor portal → Upload page. Format: question_text,difficulty (difficulty: easy / medium / hard). A sample file is at docs/question_bank/linear_regression.csv.

---

Future Scope

The current version is a working MVP. A full-scale production version would include:

1. Text-to-Speech & Speech-to-Text

Replace the text chat with a real voice interview experience. The AI interviewer speaks questions aloud (TTS via ElevenLabs or Google Cloud TTS), and students respond by speaking (STT via Whisper or Google Speech-to-Text). This makes the interview feel closer to an actual job interview and removes typing as a bottleneck for non-native speakers.

2. Camera Monitoring

Integrate webcam monitoring using computer vision (MediaPipe or AWS Rekognition) to detect:

• Eye movement and gaze direction (focus/distraction signals)
• Facial expressions (confidence, hesitation)
• Tab-switching or browser focus loss

Results feed into the report as a "behaviour score" alongside the content score.

3. Dynamic Report Generation at Scale

The current report is a single LLM call on a 5-turn transcript. At scale this would become:

• A multi-agent report pipeline — one agent scores content, one scores communication, one compares against role benchmarks
• Historical trend charts (score over multiple attempts)
• Peer percentile ranking within a batch
• Instructor analytics aggregated across hundreds of students

4. Code Editor Support

For technical interviews (DSA, system design), embed an in-browser code editor (Monaco Editor / CodeMirror) with:

• Live code execution (sandboxed via Piston API or Judge0)
• AI reviewing both the verbal explanation and the submitted code
• Support for 20+ languages with syntax highlighting and test case runner

5. Multi-Domain Scaling

The current question bank is CSV-uploaded per topic. At scale this becomes a curated, versioned question library across domains:

• Software Engineering — Java, Python, Go, DevOps, Kubernetes, AWS
• Data Science — ML, Statistics, SQL, Spark
• Management — Product Management, Business Analysis, Finance
• AI-generated question variations to prevent question leakage between students

---

Scaling to Industry Level — Cost Estimate

Component	Current (MVP)	Industry Scale (10k DAU)
Hosting	HF Spaces (free)	AWS ECS / GCP Cloud Run — ~$300–800/mo
Database	NeonDB free tier	NeonDB Business / RDS PostgreSQL — ~$200–500/mo
LLM (interviews)	Gemini free / Groq free	Gemini Flash API at scale — ~$0.075/1M tokens → ~$150–400/mo
LLM (reports)	Same	GPT-4o or Claude Sonnet for judge calls — ~$100–300/mo
TTS/STT	—	ElevenLabs + Whisper API — ~$200–600/mo
Camera AI	—	AWS Rekognition or custom CV model — ~$100–300/mo
CDN / Storage	—	CloudFront + S3 for recordings — ~$50–150/mo
Total estimate	~$0	~$1,100–3,050/mo

Key scaling decisions:

• Move from a single Docker container to microservices — interview engine, report engine, and media processing run independently and scale separately
• Use a message queue (Redis / SQS) between the interview API and the LLM workers to handle burst traffic without dropping requests
• Separate read replicas for analytics queries so instructor dashboards don't compete with live interview sessions
• Rate-limit LLM calls per student to control cost — a hard budget cap per institution is essential at B2B scale