INTELLIGENCE_REPORTS_v2.4
OpenAI Whisper (tiny) speech-to-text running 100% on-device — Web Worker inference, a WebGPU→WASM fallback chain, and zero data leaving the browser. This report breaks down the full runtime architecture.
A deep dive into the live inference pipeline: TinyFaceDetector → 68-point landmarks → 7-class emotion scoring, executed per frame on the visitor's own GPU.
How a corpus of arXiv abstracts becomes a navigable 3D point cloud — offline embedding, SVD + UMAP reduction, single-draw WebGL rendering, and in-browser semantic search over int8-quantized vectors.
Drag four corners, solve an 8-DOF perspective transform, and run Tesseract.js OCR — the entire photo-to-text pipeline stays in the browser, with no image ever leaving the device.
Embed → retrieve → generate. A full Retrieval-Augmented Generation pipeline that ingests PDFs and answers questions entirely on-device, with every source citation auditable in the UI.
How this site itself is engineered: Next.js App Router, React 19, Three.js web components, Tailwind v4 design tokens, a Vercel Blob-backed CMS, and a git-synced seed data layer.
Receive weekly intelligence reports directly to your local node.