panic-guard/gemma4-edge-agent

PanicGuard is an on‑device panic-attack triage app for iPhone + Apple Watch that detects sustained, unexplained heart-rate elevation and runs a local assessment (no cloud, no data leaves the device). When a watch-triggered haptic is acknowledged, the phone asks the user to read a short "Vocal Anchor" and combines HR features, step/activity, vocal coherence, and the user’s baseline to produce two probabilities: likelihoodPanic and likelihoodPhysicalAnomaly, plus confidence and a one-line reasoning. Who uses it: people with panic disorder (iPhone + Apple Watch users) and clinicians/researchers who need auditable, local episode logs. Key features: - Silent, passive HR detection via HealthKit - Vocal Anchor grounding + offline ASR - On-device LLM inference (Gemma4 via LiteRT) that reasons over semantic HR features - Deterministic RuleEngine mapping scores to interventions (breathing guide, grounding, emergency contact, medical alert) - Local episode logging (Core Data), exportable CSV - Implemented in SwiftUI/watchOS, WatchConnectivity, offline-first stack for privacy and low latency

AI-related keywords found in this repo's description, topics, or README summary — grouped by category. Each badge links to the corresponding ranking detail page.