Samarth Urs

AI Engineer · Conversational & Voice AI

I build real-time conversational voice AI. At Pype AI I took voice agents into production — owning Whispey, an observability platform for live voice systems, and engineering agents that automated patient interaction workflows at scale.

I'm now building fusion-runtime, an open-source self-hosted voice AI runtime that puts speech-to-text, the language model and text-to-speech in one process — ~490 ms from turn-end to first audio, with the audio never leaving your machine.

View my work Résumé

Experience

Pype AI (Singularity Corp Pvt. Ltd.)

Jan 2026 – Sep 2026 · Bengaluru

AI Intern (Jan–Jun) → AI Engineer, Conversational & Voice AI (Jul–Sep)

Built and then owned Whispey, an observability platform and voice analytics dashboard for monitoring production voice AI systems — from initial build through full-time ownership.
Engineered and maintained real-time conversational agents on LiveKit and Pipecat, automating patient interaction workflows at production scale.
Audited production call data — latency bottlenecks, call failures and conversational edge cases — to improve the reliability of live voice pipelines.

Technical Skills

Languages

Python (async/await)TypeScriptJavaJavaScriptSQLBash

Voice AI & Real-Time

LiveKitPipecatWebRTCWebSocketsSTT/LLM/TTS pipelinesBarge-in & turn detection

AI/ML Inference

faster-whisperCTranslate2llama-cpp-pythonONNX RuntimeKokoro-82MSilero VADPyTorchHugging FaceOpenCVYOLOv11

Web Development

FastAPIUvicornPydantic v2Next.jsReactNode.jsSupabase

Observability & Tooling

PrometheusStructured telemetryGitHub ActionsDockerPyPI packagingTyper CLI

Cloud & Platform

VercelGoogle CloudCloudflare PagesGit/GitHubPower BI

Projects

fusion-runtime

Self-hosted voice AI runtime · Apache-2.0 · Python

A single-process STT→LLM→TTS voice agent runtime that streams between stages instead of serialising through network hops, so synthesis begins while the LLM is still generating.

~490 ms turn-end to first audio on an RTX 3090 with a 7B GGUF model, at 127 tok/s.
Acoustic echo cancellation written from scratch — STFT filter bank, NLMS adaptive FIR, GCC-PHAT delay estimation — validated at 30 dB suppression, enabling server-side barge-in.
A runtime-adapter contract keyed on inference engine, so any Whisper checkpoint or GGUF model runs with zero new code.
Deadline-ordered, concurrency-limited scheduler sharing one loaded model across sessions, with a publicly documented degradation curve.
Shipped v0.1.0 with a Typer CLI, a browser client over AudioWorklet + WebSockets, Prometheus metrics and CI across Python 3.11–3.13.

PythonApache-2.0PyPI

View on GitHub | PyPI

FaMo (Famous Model)

React · Supabase · Vercel

A social platform built for the modelling industry, connecting talent with agencies.

Real-time backend on Supabase handling user profiles, social feeds and secure authentication.
Responsive React frontend deployed on Vercel, with a custom domain managed through Hostinger.

ReactSupabaseVercel

In development.

Earlier work

Vehicle Class Detection

A computer vision pipeline classifying vehicles in real time with OpenCV and YOLO, counting them and estimating speed from the video stream.

PythonOpenCVYOLO

Web Crawler

A Node.js link-extraction tool that walks a site, maps its internal link graph and surfaces hard-to-reach pages — useful for SEO auditing.

Node.jsJavaScript

Achievements & Certifications

Kaggle

Ranked 2271 / 13131 in "Titanic — Machine Learning from Disaster".

Open Source

Author of fusion-runtime, published to PyPI under Apache-2.0.

Infosys Springboard

Certifications in Deep Learning, NLP and Data Science.

Off the keyboard: motorcycle touring on a Himalayan 450, and learning languages.