Brief
Unsend is what happens when you build every image tool you ever needed and refuse to send a single byte to a server. All 21 tools run entirely client-side — no uploads, no API costs, no privacy tradeoff. Background removal and smart cropping use ML models that download once and run locally. Upscaling uses neural network inference in the browser. OCR extracts text with bounding box overlays. Batch processing handles queues with concurrency control. The architecture enforces strict module isolation so adding tool 24 never breaks tools 1 through 23. Fully offline-capable as a PWA.
Challenge
- Running ML inference entirely in the browser without server round-trips or model re-downloads
- Keeping 21 tool modules completely isolated — no coupling, no shared state leaking between tools
- Processing large images without blocking the UI thread
Approach
- ML models downloaded once and cached in the browser — work fully offline after first load
- Feature-Sliced Design with enforced layer boundaries so tools can never import each other
- All heavy processing offloaded to Web Workers with zero-copy data transfer
Outcome
- 21 tools shipped across 20 build sessions
- Full offline support via PWA with persistent model cache
- Zero server cost — runs in any modern browser, completely private