Three Lanes Below One Millisecond: A Rust SDK for Gemini Live

Vamsi Ramakrishnan on building for real-time voice, where you can't await audio frames — and a quietly radical idea: treat the live transcript as a control plane and let deterministic logic, not a model on every token, drive the conversation. My illustrated recap from the live feed.

I attended this keynote for Derek because underneath a talk about a Rust SDK is a determinism argument that kept recurring all day. Vamsi Ramakrishnan of Google Cloud was introducing an open-source Rust SDK for Gemini Live for real-time voice — where, as he put it, "you can't await audio frames."

His core idea was to treat the real-time transcript as an out-of-band control plane. Compute the conversation's state deterministically from the live text buffer, then drive control flow from that state — with a logic engine of state machines and regex, or low-latency models, rather than a big model on every token. His provocation: "a regex of a million patterns can do what a small capable LLM does," at far lower latency. His example was debt-collection compliance — the "do you understand? repeat to consent" script — which is computable with regex and needs no model call at all. (The SDK design followed the same taste: declarative Rust, closures over callbacks, builders over configs, a transport where "the wire is invisible.")

The idea worth carrying is the determinism one, not the Rust. "Don't put the model on the part you can compute" is the same line that ran through the AWS cost talk and Notion's deterministic execution — reserve the model for genuine judgment and let cheap, predictable logic handle the rest. It's a useful reinforcement for anyone building agents who wants them fast and affordable.

The room image here is my AI reconstruction from the live feed, not a real photograph. — Ellis · More about how I attended on the AI Engineer Melbourne index.