AI-powered mock interview platform for voice interviews, live coding, and real-time performance feedback.
Mock interviews are broken. Practicing with a friend is awkward and unreliable. Interview coaching is expensive. LeetCode doesn't simulate the pressure of speaking out loud, explaining your thought process, or being asked follow-ups in real time. Candidates go into interviews underprepared, not because they don't know the content, but because they've never actually practiced the format.
intervyu.io puts you in a real interview. You speak, and an AI interviewer named Neerja listens, responds, and adapts to your answers. She can give you coding problems to solve live in a Monaco editor with sandboxed test execution. You upload your CV and she can grill you on it. When the session ends, a performance report scores you across 5 dimensions and tells you whether you'd be hired.
Eight interview types (Google SDE, Amazon SDE, AWS/Azure/GCP Solutions Architect, CV Grilling, and more), each with a distinct persona and evaluation criteria.
Dashboard: pick your interview type and track past sessions
Live voice interview: speak naturally, Neerja responds in ~1.7s
In-interview code editor: Monaco + sandboxed execution + live test results
CV upload & analysis: upload once, Neerja references it throughout
Performance report: 5-dimension scores, HIRE/NO_HIRE, percentile benchmarks
1. Speech detection Silero VAD (neural ONNX model, Web Worker) detects true speech boundaries client-side, eliminating the false cuts you get with amplitude thresholds. Each utterance is encoded to a single WAV blob and sent over a persistent WebSocket.
2. Transcription + session fetch in parallel
The backend fires Deepgram Nova-2 STT and a session cache lookup simultaneously via asyncio.gather. STT resolves in ~500–620ms via a persistent httpx singleton (TCP+TLS reused across calls). The session is almost always already in-memory, so that lookup costs 0ms.
3. LLM streaming
Claude Haiku 4.5 via AWS bedrock-runtime converse_stream starts returning tokens in ~850ms. Conversation history is managed directly (no Bedrock Agent overhead). Neerja adapts difficulty, asks follow-ups, and drives structured interview phases defined per type.
4. Concurrent TTS
As tokens stream in, the backend splits at sentence/clause boundaries and fires Azure Cognitive Services (en-IN-NeerjaNeural, MP3) concurrently for each chunk, using a pool of 3 persistent SpeechSynthesizer instances. An asyncio.Queue sender pushes each MP3 chunk to the browser the moment it's ready.
Result: ~1.7s speech_end → first audio. ~2.9–4.0s for a full response.
Voice pipeline latency breakdown
WebSocket message sequence: full interview turn
Browser (Next.js 15)
│
│ WAV audio blob ──────────────────────────────────────┐
│ JSON control messages (speech_start, speech_end) │
│ REST calls: Authorization: Bearer <supabase_jwt> │
│ ▼
│ FastAPI · EC2 t3.small · port 8000
│ ├── WebSocket /ws/interview/{id}?token=<jwt>
│ │ ├── Deepgram Nova-2 (STT, persistent httpx)
│ │ ├── bedrock-runtime (Claude Haiku 4.5, streaming)
│ │ └── Azure Speech SDK (TTS → MP3 chunks, pool of 3)
│ └── REST /api/*
│ ├── Supabase PostgreSQL (sessions, transcripts, reports)
│ ├── S3 (CV files, audio recordings)
│ └── Lambda ×3 (code executor · cv analyzer · perf evaluator)
│
│ ◄── MP3 audio chunks (binary WebSocket frames)
│ ◄── JSON (transcript · llm_chunk · assistant_complete · coding_question)
│
├── CloudFront + S3 (static Next.js export at intervyu.io)
├── Supabase (Google OAuth + Email OTP · JWT · PostgreSQL with RLS)
└── AWS Textract (CV PDF/DOCX parsing)
| Type | Focus |
|---|---|
| Google SDE | Algorithms, data structures, system design |
| Amazon SDE | Leadership principles, coding, behavioral |
| Microsoft SDE | Problem solving, collaboration, design |
| AWS Solutions Architect | Cloud architecture, AWS best practices |
| Azure Solutions Architect | Azure services, enterprise solutions |
| GCP Solutions Architect | GCP services, data analytics |
| CV Grilling | Resume deep dive, STAR method |
| Coding Practice | Pure coding problems, optimization |
| Layer | Technologies |
|---|---|
| Frontend | Next.js 15, React 19, TypeScript, Tailwind CSS 4, Monaco Editor, Recharts |
| Backend | FastAPI (Python 3.11), WebSockets, asyncpg, asyncio |
| AI / ML | Claude Haiku 4.5 (AWS Bedrock), Deepgram Nova-2 (STT), Azure Speech en-IN-NeerjaNeural (TTS), Silero VAD (ONNX) |
| Infra | EC2, CloudFront, S3, Lambda (SAM), Textract, ACM |
| Auth / DB | Supabase (Google OAuth + Email OTP, JWT, PostgreSQL with RLS) |
- Python 3.11+, Node.js 18+
- AWS account with Bedrock, S3, Lambda, Textract access
- Supabase project
- Deepgram API key, Azure Speech key
- Create a Supabase project
- In the Supabase SQL Editor, run
database/supabase_schema.sqlto create all tables and RLS policies - Copy the project URL, anon key, service role key, JWT secret, and connection string; you'll need them in the backend
.env
cd backend
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
cp .env.example .env # fill in credentials
uvicorn app.main:app --reload
# → http://localhost:8000/docscd frontend
npm install
npm run copy-vad-assets # copies Silero VAD ONNX models
cp .env.local.example .env.local # set API URLs + Supabase keys
npm run dev
# → http://localhost:3000cd lambda-tools
sam build && sam deploy # deploy to AWS
# or for local dev:
sam build && sam local start-lambda --port 3001
# then set LAMBDA_ENDPOINT_URL=http://127.0.0.1:3001 in backend/.envMIT License. See LICENSE for details.