• Node.js
• Gemini
• Deepgram Flux
• Voice Activity Detection (VAD)

Overview

A real-time AI interview platform supporting resume-based, coding, SQL, and HR interview rounds.

• Built real-time conversational interviews using VAD, Deepgram Flux, and Gemini.
• Optimized the AI pipeline and infrastructure to reduce inference costs to ₹15–20 per interview while maintaining low-latency conversational performance.

Architecture

Overview

A low-latency, voice-first interview system that streams audio in and conversational responses out, across multiple interview formats.

Real-time voice pipeline

• Voice Activity Detection (VAD) segments the candidate's speech and decides turn boundaries.
• Deepgram Flux streams speech-to-text with low latency.
• Gemini generates the interviewer's questions and follow-ups from the transcript and interview context.

Interview formats

• Supports resume-based, coding, SQL, and HR rounds, each with its own prompting and evaluation context.

Cost & latency optimization

• The pipeline and infrastructure were tuned to hold inference cost to ₹15–20 per interview while keeping the conversation responsive — batching, streaming, and careful model/route selection at each stage.