This is an internal application behind a login, so we do not give the address. We show the mechanics, the numbers and screenshots on a call. Technically it is the most advanced project in the fleet: about 66,000 lines of code and four containers in production.
The chat has 23 tools it drives itself: creating and updating entries, reading long content in parts, attaching files, translating, searching, and pulling content from ten sources. Attachments are handled per type — an image as an image, audio and video through ffmpeg into segmented transcription, a document through text extraction with a character limit. An unreadable file produces an explicit error in the prompt and a ban on guessing its content.
A separate worker runs the source queue, fetches subtitles or transcribes with a model, classifies the content, generates the entry, retries failures and deduplicates results. Next to it stands a gateway driving a headless coding agent, with an async task API.
Model choice rests on measurement. For the chat: 288 attempts, zero errors, 4.2 s per turn and 0.0061 USD per turn — twice as cheap and twice as fast as the previous pick. For content generation: 100 and 99.4 percent fidelity across two independent runs, and 41 of 41 critical facts every time.