Taresh R.

I specialize in engineering production-grade, real-time voice agents that feel natural, not robotic. My focus is on optimizing the entire voice-to-voice pipeline including orchestration, inference, and networking to achieve end-to-end latencies under 600ms. I help companies move from "demo-ware" to scalable, production-ready voice interfaces that follow strict instructions and handle complex workflows (database lookups, call transfers, RAG) while maintaining sub-second latency. Core Competencies: Real-time Orchestration - Expert in LiveKit & Pipecat with websocket connections handling. Audio Engineering - Fine-tuning and deploying STT/TTS models (Parakeet, Canary, Whisper, Orpheus, Kokoro, KyutAI) for maximum expression and speed. Inference Optimization: High-throughput serving using vLLM and SGLang for open source models. Intelligent Turn Detection: Developed custom models to eliminate false interruptions. Open Source Work & Libraries: Conversify (100 Github stars) - A low-latency (<600ms) voice and vision enabled assistant. TTSizer (~130 Github stars) - An automated dataset creation pipeline for TTS fine-tuning . AnimeVox (>3k downloads): A high-quality TTS dataset having >11k audio clips. Orpheus Streaming: A FastAPI server achieving 160ms Time-to-First-Byte for TTS.