Intervues interview engine
A spoken real-time AI interview engine with six interviewer agent types run as phased sessions. It powers practice interviews on candidates.intervues.club and is built and run by Ayoush Chourasia, founder and sole engineer of Intervues.
The six rounds
All 6 rounds opened 3 Nov 2026, on Interview v1 on D1:
- Screening
- Deep technical grinding
- Three-stage DSA with sandboxed hidden/timed tests
- Low-level design
- High-level design on shared canvas
- AI-assisted round with a model sidekick, every prompt recorded
How it talks
Manual turn mode: each question starts and submits explicitly. Automatic endpointing was dropped. The turn budget is p95 under 4.5 seconds, guarded by a cold/warm regression harness.
The speech path runs Silero VAD with Sarvam, ElevenLabs and Deepgram for speech-to-text, Gemini on Vertex AI as the engine LLM, and ElevenLabs and Sarvam for text-to-speech. Sarvam is first in production. Grading runs through the Jev judge (jev-1.13.0), and the service runs on GCP Cloud Run in Mumbai. Recall is served by the Rust memory service.
Evidence over trust
Every cross-question and score must cite evidence: a transcript turn, a file and line, a canvas element, or a document chunk. The interviewer decides live whether to hand over a code editor or a design canvas, and every such decision logs its proposal, confidence, rationale, and veto.
Grading is judge-model based: one judgment per question, three-band confidence gating, and a deterministic fallback whenever the judge is unsure or the breaker is open. A contradictory-candidate fixture tests whether the interviewer catches self-contradiction.
Status
MRR: Not yet reported.