Real time is the product.

A synthetic presenter that reports to senior management, generated as the meeting happens.

§ 01 — Brief

A major cloud corporation in the Middle East needed avatars that speak to people — senior management — and report business information in real time. Not a clip rendered in advance and played back: a synthetic presenter that generates and speaks as the conversation happens.

§ 02 — The constraint

Real time is the whole problem. A briefing avatar is measured against human patience, not against a render farm’s schedule. Cross the latency budget and the presenter stops being a presenter — it becomes a slow video, and the person in the room stops listening. Every other decision is subordinate to that budget.

§ 03 — Optimization is the work

The budget was met on the GPU. We optimized the generation path where the cost actually accrued, and added session balancing so that as concurrent presenters multiplied, each kept correct, intact speech instead of degrading under load. The reusable core of that speech work was factored into a dedicated speech-optimization library written in Rust.

§ 04 — Result

The solution was deployed.