Porting Pipecat to Rust: The story of Rustvani-Fast Cheap Voice Ai for Bharat

Voice AI is having its infrastructure moment. But if we look closely at the stack everyone recommends: For eg: Pipecat with 12k git stars is getting popular. but every pipecat layer assumes you have AWS credits, and Series A funding. For the vast majority of Indian builders who are cash-strapped startups, solo developers, govtech contractors working on ₹-denominated contracts, this stack simply burns. Due to startup time they have to run warm instances, that too atleast 512mb perpetually.

Rustvani is a direct response to this mismatch. It is a Rust-native voice AI framework inspired by Pipecat's pipeline architecture, rebuilt from scratch to run on cheap infrastructure, serve Indian languages using Indian providers, and fit inside a 30MB static binary.(pipecat takes 400mb and 2 sec start). which means it can run 3 instances comfortably on a 128 mb flyio instance.

This talk is the story of that build and the decisions, the dead ends, and the design principles that emerged. I'll walk through why I started with Python Pipecat, what broke when I tried to deploy it under real Indian cost constraints, and why Rust became the answer. Eventhough the title says porting pipecat to rust, Rustvani can do much more. Along the way I'll cover: replacing the ONNX Runtime dependency with a pure Rust Silero VAD implementation, building RAVI (a lightweight inter-component protocol), the Dhara conversation flow engine (a node-based state machine for managing dialogue), and integrating Sarvam AI for Indian language STT and TTS. The result deploys as a single static musl binary to a Fly.io instance in the Mumbai region. Absolutely no GPU, no Python runtime, no cloud vendor lock-in.Also a 300ms cold start on fly io (of which 200ms is firecracker startup)

The deeper argument of this talk is about who this FOSS infrastructure is designed for. Most voice AI tooling is built by well-funded San Francisco companies, for well-funded San Francisco companies. Rustvani is an attempt to build from the other direction ie, starting with the constraints of an Indian indie builder and working outward. I think that inversion produces better software for most of the world.

1)Why the current voice AI tooling stack has a structural bias toward well-resourced, US-centric infrastructure and what that costs Indian builders?

2)How to architect a real-time voice pipeline in Rust without GPU dependencies or heavy runtime overhead. Right to not keep a warm instance.

3)What "India-first" constraints (cost, latency, vernacular languages, data residency) actually look like when translated into systems design decisions.

4)How a solo FOSS builder can take a published Python framework, port it to Rust, and ship it to crates.io and what that journey really looks like.?

5)Future of Rustvani where rustvani clients will be run on ESP32 boards and rustvani will be part of unified front-end back-end and voice ai framework- by blending Rustvani and Dioxus.