Quietcare

An ambient AI safety companion for elderly care. Cal Hacks 2026 · June 20–21

"Your phone doesn't blow up every time grandpa drops a spoon — only when something is actually wrong."

Elderly people living alone face a dangerous gap: when something goes wrong — a fall, a stroke, choking, sudden confusion — help often arrives far too late. Quietcare watches ambiently, runs a calm spoken check-in when something looks off, and loops in the right person fast, without false alarms.

How it works

A phone app watches motion + audio cheaply on-device. It continuously streams short audio segments as audio_probe frames — the backend tags the YAMNet audio scene and watches for a wake signal — and the elder can hold to talk to send a voice_conversation clip that the backend transcribes, answers with the conversation agent, and speaks back. When something looks wrong it streams a trigger event over the same WebSocket. The backend then runs two cooperating Claude agents:

• elder-agent — reasons about the event, runs a spoken voice check-in, fuses the signals (trigger source + transcript + whether the elder responded + any non-speech acoustic distress), and decides to resolve or escalate.
• caretaker-agent — triages an escalation off the message bus and alerts the human caretaker via SMS / voice call, and can request human-authorized 911.

 phone (client)              cloud (server)                          humans
 ┌──────────────┐  WS /ws   ┌───────────────┐   BAND bus  ┌────────────────┐
 │ fall detect  │ ───────►  │  elder-agent  │ ──@mention─►│ caretaker-agent│
 │ audio buffer │  trigger  │  (Claude)     │             │   (Claude)     │
 │ check-in I/O │ ◄───────  │  voice + ML   │             │  Twilio SMS/call│
 └──────────────┘  speak/   └───────────────┘             └───────┬────────┘
                   listen                                          │ gated
                                                                   ▼
                                                        human-authorized 911

Everything runs end-to-end with zero credentials — every external provider sits behind an interface with an automatic mock fallback.

Repository layout

Quickstart

Backend (all-mock, no keys)

cd server
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env            # all keys may stay blank

# terminal 1 — run the server
uvicorn app.main:app --host 0.0.0.0 --port 8080

# terminal 2 — drive the full loop
python mock/send_trigger.py --scenario emergency   # -> escalation + caretaker alert
python mock/send_trigger.py --scenario fine        # -> resolves silently

GET http://localhost:8080/health reports which provider impls are active (mock vs real).

Client (device app)

See client/README.md for the Expo dev-build + Android install steps. The client holds no API keys and talks to the backend purely over the WebSocket contract.

Safety model

The LLM decides transitions (via which tools it calls); the code enforces them and the hard invariants:

• 911 is never reachable without explicit human confirmation. escalate_911 is gated in state_machine.gate_911(human_confirmed=…), re-sanctioned by a policy gate, and dispatch only fires on a token-verified human approval.
• Every escalation passes through a check-in first, unless the trigger is an unambiguous hard fall.
• Agents never autonomously dial emergency services — they can only request human authorization.

idle -> triggered -> checking_in -> (resolved | escalating)
                  \-> escalating              (only if hard_fall)
escalating -> caretaker_notified -> (human_ack | 911_gated)

Sponsor / integration stack

Every integration is wired behind an interface with an automatic mock fallback.

Secrets live only in server/.env / client/.env (gitignored) and the BAND daemon agent_config.yaml. Never commit keys. See server/.env.example for the full list.

Verification scripts

The backend ships live diagnostics under server/scripts/:

• check_keys.py / check_remaining_sponsors.py — live connectivity for every provider.
• verify_audio.py / verify_yamnet.py — Deepgram STT↔TTS round-trip + YAMNet classification.
• verify_app_hook.py / mesh_demo_kickoff.py — drive the app→elder→caretaker BAND mesh end-to-end.
• verify_emergency_call.py — places a real gated emergency-dispatch call (to a number you control).
• verify_observability.py / probe_armoriq.py — Arize tracing + ArmorIQ security-scan checks.
• fake_phone.py — simulates the phone over the WebSocket, sending a real spoken clip through the voice_conversation STT→agent→TTS loop.
• fetch_yamnet.py — downloads the YAMNet model into server/models/.

Run the test suite with python -m unittest discover -s tests from server/.