Meta — Codec Avatars

The challenge

Codec Avatars — Meta's photorealistic telepresence research — needed quantitative and qualitative evaluation of the avatar pipeline at research pace. Evaluators needed tools that didn't exist, and the evaluation process itself was the bottleneck between researchers and a better avatar.

The decisions — and why

Decision: Build the evaluators' tools in both VR and 2D (Unity/C#), rather than a single surface. Why: Defect detection is medium-dependent — some avatar defects only read in VR, others faster on flat screens. Tradeoff accepted: [PLACEHOLDER — authoring session]

Decision: Rebuild the team's components as event-driven, decoupled infrastructure before scaling the evaluation process. Why: A highly-interdependent component set can't be scaled or threaded; independence made the restructure possible at all. Tradeoff accepted: [PLACEHOLDER — authoring session]

Decision: Co-lead a PyTorch machine-learning initiative to automate quality evaluation. Why: Human evaluation throughput has a ceiling; automated testing scales past it. Tradeoff accepted: [PLACEHOLDER — authoring session]

The outcome — with numbers

• 2 hours saved per evaluation — the process restructure, in production
• 200% increase in evaluation scaling capacity
• Automated data-collection tools (in-engine + native C++ plugins) delivering debugging information directly to researchers