Hearthwatch

Hearthwatch is an AI companion for families looking after an aging parent at home. A single camera watches over Mom or Dad; computer vision tracks daily activities; Claude reasons about what's happening and only calls the family when something actually matters in the language a daughter or son would actually use rather than bombarding their phones with event notifications. Most elder-care monitoring tools detect events and beep. Hearthwatch understands situations and talks. When a fall happens, it doesn't fire a generic alarm — it looks at the saved keyframe via Claude Vision, confirms what it sees, and composes a warm alert: "Dad may have just fallen in the living room near the coffee table — call him now, and if he doesn't answer, head over right away." That specificity comes from Claude actually reading the frame. The system runs two layers in parallel. Local detectors (MediaPipe pose + YOLOv8) handle every frame on-device fall detection (including a close-to-camera vanish path for when someone falls below the field of view), pacing, inactivity, repetitive behavior, and multi-person subject tracking that follows the right person even when others enter the frame. Claude runs on a tiered cadence: Vision on keyframes to confirm and describe, Opus to reason over the rolling event window and assess risk, Haiku to compose alerts. Every answer in the chat tab is grounded in a real SQLite query — no hallucination. It ships on a Raspberry Pi 5 with a Pi Camera Module 3, costs under $150 in hardware, and runs at roughly $0.23 per monitoring session in API credits. The goal was to build the thing a real family would actually put on their kitchen counter.