SamarthUrs18

Samarth Urs is an AI engineer and solo developer building real-time conversational voice AI. His current open-source project, fusion-runtime, runs speech-to-text, an LLM, and text-to-speech in one self-hosted Python process so audio can remain on the user's machine and responses can stream with low latency.

Visit Website

At a Glance

BengaluruHeadquarters

AI Tools by SamarthUrs18

(1)

fusion-runtime

Self Hosted Voice Agent Runtime

Voice Assistant Autonomous Systems Local Inference

Discussions

No discussions yet

Be the first to start a discussion about SamarthUrs18

Latest News

09/26/2026

fusion-runtime v0.1.1 released with tool calling and measured multi-caller GPU performance

github.com

09/26/2026

Project documentation updated with vLLM/SGLang concurrency results: about twelve callers answered in about a second on one RTX 3090

fusion-runtime.dev

09/26/2026

PyPI package summary revised to clarify the approximately 490 ms post-turn processing figure versus approximately 991 ms from the caller's last syllable

github.com

09/21/2026

fusion-runtime v0.1.0 released as an Apache-2.0 self-hosted voice-agent runtime

github.com

Products & Services

fusion-runtime

2026-09-21 (v0.1.0); 2026-09-26 (v0.1.1)

Apache-2.0 Python runtime and PyPI package for self-hosted voice agents. It runs speech-to-text (faster-whisper), an LLM (llama.cpp or external vLLM/SGLang/llama-server-compatible endpoints), and text-to-speech (Kokoro) on hardware controlled by the user, streaming between stages so speech can begin while generation continues. v0.1.1 adds Python-function tool calling with spoken preambles, cancellation on interruption, remembered tool results, and multi-caller serving support.

Market Position

Positioned as a local, self-hosted alternative to multi-vendor cloud voice stacks such as Deepgram for STT, OpenAI/Groq for LLM inference, and ElevenLabs/Cartesia for TTS. Its differentiation is control over latency, data locality, deployment, and marginal cost; its stated trade-off is that open-weight models currently have a lower quality ceiling than the best closed APIs, and telephony and very large-scale concurrency are not yet supported.

Leadership

Founders

Samarth Urs

AI Engineer focused on conversational and voice AI. He studied B.E. Artificial Intelligence & Data Science at CMR Institute of Technology (2022–2026), worked at Pype AI (Singularity Corp Pvt. Ltd.) in Bengaluru from January to September 2026 as an AI Intern and then AI Engineer, and built/owned the Whispey voice-AI observability platform and production voice agents using LiveKit and Pipecat. He is the author of fusion-runtime.

Founding Story

Samarth started fusion-runtime to provide a self-hosted alternative to assembling voice agents from multiple metered cloud APIs. The project's initial vision was to keep STT, LLM, and TTS on hardware controlled by the user, avoid per-minute API bills and network hops, stream the stages into one another, and support privacy-sensitive and local-first deployments.

Business Model

Revenue Model

fusion-runtime is distributed as open-source software under Apache-2.0 and is installed from PyPI; the project emphasizes zero per-minute API cost when run on user-controlled hardware. No subscription or usage pricing is published. A hosted cloud version is described as planned but not yet built.

Target Markets

Industries & Segments

Developers and teams building self-hosted voice agents
Privacy- and data-residency-sensitive organizations
Narrow-domain customer-service and operations workflows
Cost-sensitive, high-volume voice applications
Local-first developers and embedded/device applications

Use Cases

Narrow-domain voice agents such as reservations, order-status and logistics updates
IVR replacement and other conversational voice workflows
Privacy- and data-residency-sensitive teams that cannot send audio to a third-party cloud
Cost-sensitive, high-volume deployments where per-minute API metering is undesirable
Local-first applications including kiosks, robots, desktop apps, and single-board devices
Embedding a voice agent into a website through the supplied browser client

Quick Facts

Headquarters

Bengaluru

History & Milestones

2026

Samarth Urs began building fusion-runtime as an open-source, single-process self-hosted voice-agent runtime; the project was published under Apache-2.0 and packaged for PyPI.

2026-09-21

Released fusion-runtime v0.1.0 with the core STT→LLM→TTS runtime, Typer CLI, browser client over AudioWorklet/WebSockets, Prometheus metrics, and Python 3.11–3.13 CI. The release documented roughly 490 ms turn-end-to-first-audio processing on an RTX 3090 profile.

2026-09-24

Published v0.1.1, adding tool calling, named vLLM/SGLang/llama-server backends, expanded model/runtime settings, and clearer model/deployment errors.

2026-09-26

Published concurrency measurements showing that vLLM and SGLang enabled approximately twelve callers to receive responses in about a second on one RTX 3090, while speech stages became the next bottleneck.

2026-09-26

Updated the PyPI summary and project documentation to distinguish approximately 490 ms processing after turn-end from approximately 991 ms measured from the caller's last syllable, including the configurable silence wait.

Key Capabilities

Single-process, local-first STT→LLM→TTS streaming

faster-whisper speech-to-text, llama.cpp/vLLM/SGLang/llama-server LLM backends, and Kokoro text-to-speech

Voice activity/turn detection and barge-in interruption that stops generation and playback when the caller talks over the agent

Browser client served by the runtime, embeddable with two script tags and using AudioWorklet/WebSockets

Short-lived browser session tokens, bearer-key authentication, origin controls, concurrency caps, and operational limits

Python-function tools with docstring/type-hint schemas, timeout handling, cancellation, and tool results retained in conversation context

Integrations & Partnerships

Platform Integrations

Python package installable from PyPI (fusion-runtime)
llama.cpp for in-process inference
vLLM, SGLang, and llama-server through model-server adapters/endpoints
Hugging Face model references and local model files
Browser embedding through fusion-runtime.js, WebSockets, and AudioWorklet
Prometheus metrics

Connect

Website

fusion-runtime.dev

GitHub

SamarthUrs18

AI Topics

SamarthUrs18 focuses on these topics:

Voice Assistant(1)

Autonomous Systems(1)

Local Inference(1)

Back to all developers Suggest an edit

SamarthUrs18