Skip to content

Research Hub

Decision logs, technical deep dives, and architectural findings captured throughout development.

Topics

  • Thesis Scope

    Overview of the thesis, comparison criteria (load time, inference speed, memory, ease of integration), target browsers, and initial code review findings.

    Read

  • Runtime Selection

    Analysis of 3 runtimes (ORT Web, TF.js, LiteRT.js) and 3 frameworks (Transformers.js, MediaPipe Tasks, ml5.js). Backend matrix and model distribution plan.

    Read

  • Technical Deep Dives

    ONNX model sourcing (HuggingFace CDN), CPU vs WASM distinction in TF.js, WebNN support, MODEL_REGISTRY architecture, and runtime requirements (SharedArrayBuffer, COOP/COEP headers).

    Read

  • Model & Task Distribution

    Full task catalog: 16 tasks across vision, audio, and NLP. Top 5 models + datasets per task, runtime compatibility matrix, and core benchmark task selection.

    Read

  • Benchmarking Plan

    Three-dimensional benchmarking methodology: runtime comparison, framework comparison, and abstraction overhead analysis.

    Read

  • NPU-WebNN Research

    WebNN API research and NPU benchmarking approach.

    Read

  • Storage Architecture

    How benchy/ persists data across the four browser storage mechanisms, why each mechanism was chosen, and the known gaps.

    Read