From ideas to code: testing attention theory in practice
I built Mindstream as an engineering experiment and a working app: mindstream.app.wiredgeese.com. It reads Habr via RSS, generates LLM annotations and LLM digests, and assembles a single feed. Personalization happens in the browser: user actions are interpreted as attention signals, and a local interest vector is built from those signals.
The interest score is computed from the digest embedding: the digest serves as a compact carrier of the post’s meaning, the embedding places it in semantic space, and the distance to the current interest vector is used as a relative indicator of correlation. I wrote the app using the ADSM (Agent-Driven Software Management) approach: documentation captures invariants, code follows from it, and agents are used as a tool for guided development. The full article in Russian is published on Habr.