AI · Real-Time· 2024 · Full-stack engineer
Real-Time AI Interview
WebRTC interview simulator that streams candidate audio to GPT evaluators for live feedback, scoring, and transcripts.
- Reduced interviewer prep time by 40%
- Under 300ms round-trip latency for feedback
Next.jsNode/WebSocketRedisOpenAI RealtimePostgreSQL
Context & Goals
Interview coaching platforms often lack live, contextual feedback. This project delivers a browser-based interviewer that listens, scores, and surfaces coaching tips in real time without sacrificing privacy or stability.
Architecture Overview
- Client: Next.js front-end with WebRTC publisher, noise gating, and timeline UI.
- Signaling: Node server establishing WebRTC sessions, handling TURN/STUN negotiation, and streaming audio frames to evaluators.
- Inference layer: GPT-based evaluators running in parallel, enriching transcripts with rubric checks (communication, technical depth, empathy).
- Storage: Redis for transient session data, PostgreSQL for final reports, highlights, and improvement suggestions.
Highlights
- Built adaptive scoring that adjusts rubrics mid-session based on candidate responses and job role.
- Implemented safe-words and privacy controls; sessions can redact sensitive topics before saving transcripts.
- Added playback controls that pair video segments with constructive coaching cards.
Next Steps
- Expand to support collaborative interviewer panels with shared annotations.
- Integrate HRIS exports to close the loop on performance tracking.
Up next
Fault-Tolerant Federated Learning
Federated learning approach for sensor systems that handles missing data and sensor dropouts using imputation and LLM-based filling, evaluated on the MHEALTH dataset.