Sift

Sift is an AI performance engineer for Three.js applications. Developers connect a project, describe what feels slow, and watch agents investigate its source code, build an optimized candidate, and verify the result beside the running 3D scene. Meta powers the Investigator and Optimizer. Subquadratic assesses regression risks and selects additional checks. Fetch uAgents coordinate revision-checked handoffs. An independent deterministic verifier runs builds, visual comparisons, object-transform checks, interaction probes and performance benchmarks. Failed candidates are rejected; accepted changes come with evidence and a downloadable patch. In our tested 2,000-robot simulator, Sift reduced draw calls from 6,009 to 12 and CPU submission p95 from 19.6 to 1.8 milliseconds. All 16 checkpoints and 80 picking probes passed, and we manually verified robot selection and dispatch. Sift also includes a bounded Skill Lab that learns versioned, reusable optimization transformations and adopts them only after evaluation on held-out applications. Together, these workflows address Trust but Verify and The System That Gets Better: measurable improvements, independent acceptance decisions, and reusable skills supported by executed evidence.