TempoAI

Tempo finds the moments that matter in whatever you are doing, and helps right then. Try it live: https://tempo-dev.k8s1.aviproduction.org - this is the same deployment shown in the demo video. It opens straight into a shared demo account, no sign-up needed (the first load takes about 20-30 seconds while the session warms up). In the browser you can explore the planner and the Tempo panel; live camera/screen capture with spoken tips runs in the Tempo Mac/iPhone app, as shown in the video. Most AI assistants wait to be asked. Tempo watches the activity you choose to share (camera or screen) and listens to your mic. When something important happens, such as a step done wrong, a stall, a question aimed at it, or a change worth noticing, it speaks up with a short hint by voice or text. It works across processes: not one scripted scenario, but "is this moment important in the process I can see?" How it works: 1. Capture: the Tempo Mac/iPhone app streams camera or screen plus mic through LiveKit into media-ingest. 2. Perception: media-perception runs a cheap CPU cascade (motion-compensated frame change, then the MobileCLIP2 encoder for novelty and dedup, then a per-person token-budgeted moment scheduler). Only rare frames reach a judge model, which returns process, stage, importance, kind, why and confidence. Speech goes VAD, then streaming transcription, then a check that the question is addressed to Tempo. 3. Moments: important events are published as moment.detected. 4. Session: session-coordinator puts moments into a Jarvis session (one working session per capture). A lead agent decides what to say. 5. Answer: the reply goes back to the person as text and speech, played through LiveKit. A session summary and capture history are kept at the end. Built during the hackathon (Oct 26 - Nov 4): - New services: media-ingest, media-perception and session-coordinator (repos created Oct 31 and Nov 4). - Tempo Mac and iPhone apps (Tauri builds), live camera/screen capture and switching, voice playback. - Changes to existing services: streaming speech transcription, speech.speak and a Meta fallback in runtime; perception.screener and perception.strong model classes plus a Meta ASR route in model-gateway; scene triggers and one-session-per-capture in flow-engine; capture relay and a generic CapabilityService in jarvis-adapter; capture routes in the gateways. Existed before the hackathon (not claimed as hackathon work): the Jarvis platform (agent sessions with a lead agent, memory "brain", runtime, IAM, terminal, cockpit, flow-engine, model routing) and the Tempo planner app (since Sep 16, renamed to Tempo on Oct 29). Not yet done: Meta glasses and watch as sources are the next step and are not shown in the demo.