Akash Singh

Designed and shipped a real-time voice-to-voice agent over WebRTC, integrating streaming ASR LLM reasoning TTS into a full-duplex production pipeline for travel booking. Architected a multi-agent conversational AI system (LangGraph StateGraph, GPT-4) coordinating 10+ agents for search, filtering, and booking; achieved 95% intent recognition accuracy. Integrated live travel APIs via async clients with circuit breaker patterns; maintained 99.5% system uptime in production. Led hallucination mitigation through prompt engineering, session feedback, and synthetic evaluation.

Led end-to-end TTS research across Tacotron, FastSpeech, and HiFi-GAN, covered single-speaker, multi-speaker, and multilingual settings across 11 Indian languages at 5M calls/day. Built and deployed streaming ASR systems (DeepSpeech, Whisper, Kaldi), developed full NLU pipeline from data creation to cloud deployment on Azure and AWS. Distilled a large recommendation model into a compact on-device model deployed in production on Android inside a keyboard product for real-time content recommendation. Led a cross-functional team of engineers, linguists, and CUX designers across the full research-to-deployment lifecycle.

Trained ELMo and ULMFiT language models from scratch in 9 Indian languages; applied to entity tagging, text classification, semantic role labelling, and POS tagging. Built speaker recognition pipelines (X-vector, D-vector) achieving 95%+ accuracy for phrase and non-phrase tasks; developed keyword spotting with TensorFlow.js for browser deployment. Built dialog policy via deep RL; developed keyword spotting with TensorFlow.js for browser deployment.