UnThink

[Framing]

The architectural decision that shaped everything else was the dual backend. Django handles auth, quota, and the save API, with under 300ms return so the browser extension feels snappy. FastAPI handles the AI classification pipeline and the SSE stream that updates the dashboard when classification completes. They talk via Redis pub/sub and Celery (three queues, classification, scraping, and reports) and never call each other directly. This split exists because Django is great for relational CRUD and FastAPI is great for streaming LLM responses. Using one for both was tempting but wrong.

[Build]

SSE over WebSocket was an early call. We don't need bidirectional, and SSE just works through any HTTP proxy. The browser extension (Manifest V3, Chromium only) is the primary capture interface. Right-click on selected text, hit Send to UnThink, done. The folder taxonomy is the part I'm proudest of. Every fragment lands in exactly two folders, a topic folder like Tech/DevOps/Redis and a source folder from the page title. Tags equal folder membership, so editing a tag later is just moving the fragment between folders.

[Deploy]

Quota limits are honest. 25 right-click saves a day, 250 a month. I'd rather have a clear ceiling than a vague fair-use line. We're still in development, but the capture, classify, and file loop is solid.

[What's next]

What's next. Replacing the Jina and Firecrawl fallback scrapers with a single in-house crawler, plus an offline mode for the extension.