The pitch

Every fitness band, phone, and smartwatch already collects a rich stream of signals — heart rate, HRV, sleep stages, motion, location, blood oxygen. Today that data gets funneled into a vendor’s cloud, and whatever “insights” come back are shallow, ad-adjacent, and locked behind a subscription. I want the opposite: a DIY assistant I run myself that ingests my own sensor streams, reasons over them locally, and nudges me in real time — “your HRV tanked, maybe skip the hard workout,” “you’ve been sedentary for 3 hours,” “you sleep worse the nights you eat after 9pm” — delivered through the interfaces I already use: Siri and Apple push notifications.

The non-negotiable is data ownership. Sensor data lands on hardware I control (the homelab Mac Mini is the obvious host), inference runs against a local model, and nothing about my body is shipped to a third party. It’s the self-hosted, privacy-first answer to Whoop/Oura/Fitbit coaching — an assistant that happens to live in my house instead of theirs.

Rough shape: a HealthKit / wearable bridge that syncs readings to a local store, a rules-plus-LLM layer that decides when something is worth surfacing, and a thin delivery path out to Siri (App Intents / Shortcuts) and APNs for the actual feedback. Open source so anyone can point it at their own stack.

Open questions

• Ingest path. HealthKit export is easy but not real-time; getting live wearable data off-device (Apple Watch especially) is the hard part. How close to realtime can this actually get without a first-party watch app?
• Siri surface. App Intents vs. Shortcuts automations vs. a full companion app — what’s the lightest way to make it conversational both directions?
• The reasoning layer. How much is deterministic rules (“HRV < baseline → alert”) vs. an LLM summarizing trends? Where’s the line where a local model earns its keep over a few thresholds?
• Push without a backend. Can APNs be driven cleanly from a self-hosted box, or does it need a tiny always-on relay? (Ties into the homelab setup.)