HALO - Second Brain
A second opinion that never leaves your phone. HALO connects your hospital records to your smartwatch and turns the stream into plain-language health insight. Every model runs on the Snapdragon NPU through ExecuTorch, nothing is uploaded to cloud, ever.
Core trick
Pulls Health Connect data and generates per-organ summaries with an ExecuTorch-compiled LLM on the NPU, with no uploads.
Tags
Description
$25
README excerpt
# Health Second Brain — Android (Hackathon Build) Native Android app for the **ExecuTorch Hackathon (Qualcomm × Meta, SF, Jun 27–28 2026)**. Target device: **Samsung Galaxy S25 Ultra** (Snapdragon 8 Elite, SM8750). ## Stack - Kotlin + Jetpack Compose UI - Health Connect (`androidx.health.connect:connect-client`) for on-device health data - ExecuTorch (NPU-backed Llama / Phi via QNN) for the three LLM-generated strings per organ — wired through `OnDeviceLlmService` (currently a stub returning canned strings) ## Screens | Route | File | |---|---| | Home (organ bubble cluster, dark) | `ui/screens/HomeScreen.kt` | | Scan focus (rings + preview card) | `ui/screens/ScanFocusScreen.kt` | | Organ detail (metrics, charts, sentiment, next step, chat entry) | `ui/screens/OrganDetailScreen.kt` | | Full chat | `ui/screens/ChatScreen.kt` | ## Design notes - Built from `DESIGN_SPEC.md`, but the **palette is custom** and every surface is **dark mode** (the spec switched detail screens to light — we don't). OLED-friendly, judging-friendly demo on the S25 Ultra display. - Caveat/Gaegu are not bundled — type system uses Serif/SansSerif as drop-in stand-ins to keep first-build deterministic. Swap to Downloadable Fonts when wifi is reliable. ## Build 1. Open in Android Studio Koala+ (AGP 8.5). 2. Generate the Gradle wrapper once: `gradle wrapper` (or let Studio do it on first sync). 3. Plug in the S25 Ultra → Run. ## ExecuTorch integration plan 1. Build (or download) a `.pte` for Llama 3.2