Hassan Kazi

Lead the AI & real-time stack for InCBot (AI video interviewer) and IncScreen (AI phone & WhatsApp screening, which I built from scratch on Twilio and Exotel telephony with Vapi); scaled to 1,000+ concurrent interviews. Root-caused why multilingual interviews broke: English-only text guards failed on Devanagari (0.39 vs 0.5 match threshold) and Marathi STT hallucinated in 10 of 15 turns; drove script-aware, feature-flagged fixes.

Worked on WhatsApp automation and real-time voice agents. Full-stack MERN development (MongoDB, Express.js, React, Node.js) across multiple projects, including Raise Wademo, Hobbit and Ping Me: frontend implementation, backend API integration, database operations, debugging and feature development.

Shipped 5+ responsive React/JavaScript apps for 4+ clients, cutting page load time by 40%. Refactored 3 legacy codebases, raising SonarQube maintainability scores by 35%.

Built a 5,000+ rule Boston parking-regulations dataset for autonomous-vehicle rule validation. Designed a real-time data pipeline and frontend used across 10+ AV models, with a US team.

My own Hindi/Marathi ASR and TTS models and a LoRA-tuned LLM, joined by a from-scratch streaming orchestrator (VAD, end-of-turn, barge-in); ONNX, INT8, speculative decoding.

Target: ~800 ms p50 voice-to-voice, ~5× lower cost/min than hosted APIs.

STT → LLM (tool calling) → TTS interviewer with feature-flagged guards; Docker CI/CD to LiveKit Cloud via GitHub Actions; runs in 42+ languages.

Built from scratch: AI screening phone calls over Twilio and Exotel telephony, orchestrated with Vapi, plus WhatsApp interviews. Multilingual question sets, 7 Hindi voices, voicemail detection and fluency scoring. Hindi call-quality work: replies took ~5 s vs ~2 s in English, plus cut-off speech and interruptions.

Talks in English, Hindi or Marathi and switches when you do. Three-deep fallback chains for STT, LLM and TTS; keyterm-prompted STT; per-reply voice routing (Hindi and Marathi share a script, so it reads the reply's words); false-interruption recovery; Krisp noise cancellation; hangs up on silent callers. Five Python files, deployed to LiveKit Cloud by GitHub Actions.

Opus VBR made macOS show a 3:32 recording as 6:53+. Adds bit_rate / constant_bit_rate so recordings can match Egress (constant 128 kb/s).

A failed encoder setup made AgentSession.aclose() hang forever and queued audio in memory for the rest of the session.

A closing tag without its opener drove a depth counter negative and split <emotion> spans across TTS chunks. One-line fix.