What I learned building JARVIS’s daily voice briefing
JARVIS is my personal AI chief-of-staff — a voice-first HUD that’s supposed to tell me what actually matters about my day without me having to go looking for it. Most of this stretch was about making that briefing trustworthy rather than just present.
The email triage loop was re-judging the same roughly 20 threads every five minutes, all day, for no reason — burning real work for zero new information. Caching a verdict per thread, keyed on the thread id and its last message, cut that down to about 7 genuine judgements a day; identical polls now cost nothing.
The voice digest itself had a quieter bug: it was reading a field that had never actually existed in the data model, so it stayed blind to the real inbox and the real attention queue no matter what was actually happening. Fixed, it now carries the real blockers plus which integrations are actually down — which matters more than almost anything else it says.
DESK, JARVIS’s notes surface, grew a Trello-style kanban board and cards that pin straight to an Orbit project. And it got a new "Hex Core" icon, legible even at 16px, which sounds cosmetic until you’re staring at a blurry favicon fifty times a day.
None of this is glamorous. It’s the unglamorous work of making an assistant say true things reliably, which turns out to be the entire point of having one.
