CalmCue

Inspiration

Could physiological signals be context-aware?

I started with a much more ambitious vision: a privacy-first assistant that continuously unifies wearables + calendar + email to detect stress and coach you in real time. Very quickly I learnt that production-grade permissioning (OAuth scopes, SDK approvals, data-use reviews) would take longer than a weekend. So we pivoted to build a focused, working slice of that vision—proof that the core loop works end-to-end. This submission is a snapshot of what’s possible.

What it does

Watch the video. CalmCue listens to your heart-rate stream and pulls your calendar context. When it detects pre-event stress patterns, it triggers a lightweight coaching nudge or a joke, or just plain makes fun of you (e.g., paced-breathing cue), then logs whether your physiology settles down.

How we built it

• Stack: Next.js (Vercel) dashboard + Supabase (Postgres) + Edge Functions (Deno) + Python workers + WebSockets.

• Data sources:

  • Google Calendar → ingest-gcal Supabase Edge Function (handles recurring events, time zones).
  • Wearable HR (HRM Pro Plus) → local Python DB via sqlite3 → WebSocket → Supabase physio_measurements.

• Storage model (layered):

  • Layer −1: sources (where data came from).
  • Layer 0: events, physio_measurements.
  • Layer 1: inferences (instant reasoning on rolling windows; multiday trend jobs).
  • Layer 2: features (e.g., pre-meeting HR slope, baseline deltas) feeding the UI.

• Reasoning: Cohere-backed prompts combine current signals + recent trends + upcoming events to decide when to nudge.

• UI: shadcn/Tailwind + Recharts for live graphs; Supabase Auth for session gating.

• Ops: Cron-style Edge Function for slower multiday analysis; instant inference runs on insert triggers / websocket events.

Challenges we ran into

• Our Supabase schema got too convoluted at first—hard to query and reason about. We simplified to the “Layer −1 → 2” pattern above.
• Calendar data is messy: recurring events and far-future instances polluted tables; we added filters and cleanup.
• WebSocket reliability + noisy HR required smoothing windows and reconnection handling.
• Typical hackathon pain: OAuth scopes/rate limits and time-zone edge cases.

Accomplishments that we're proud of

• A working end-to-end loop: ingest → infer → coach → observe effect.
• Live HR streaming to the dashboard with event-aware triggers.
• Clear, documented schema and jobs for both instant and multiday reasoning.
• Privacy-first defaults and a demo that still feels useful.

What we learned

• Start from the smallest valuable loop, then layer sophistication.
• Good feature engineering (slopes, baselines, context flags) beats fragile heuristics.
• Keep schemas boring; make the read path easy.
• “Ambition vs. permissioning” is real—design for local-first demos while planning for compliant integrations later.

What’s next for CalmCue

• Proper permissioned integrations (Garmin/HealthKit/Outlook/Gmail) and background agents for mobile.
• More signals: HRV and accelerometer; better personalization of baselines.
• A unified stress score + evaluation dataset to measure real impact.
• Push notifications + voice guidance, and on-device processing for stronger privacy.
• Production-grade schema (fewer tables, clearer contracts) and a public API so others can build on the core loop.

Built With

• auth
• bluetooth
• cohere-api
• google-calendar-api-(oauth-2.0)
• next.js-(react)
• python
• recharts
• shadcn/ui
• storage)
• supabase
• supabase-(postgres
• supabase-edge-functions-(deno)
• tailwind-css
• typescript
• vercel
• websockets