WristMind (ai.wristmind.app) — 4 GB Android 15 Watch Edge-First AI Showcase

Rearchitecture of the legacy cross-platform Flutter OpenMMPA assistant into a zero-Cloud-Functions, edge-first native Kotlin Android application engineered specifically for a Full Android 15 rectangular watch (>2" display, Smallest Width = 426dp, 4 GB physical LPDDR RAM, Google Play Services).

---

Documentation

• Architecture Specification: In-depth technical architecture covering the 4 GB LPDDR memory topology, LiteRT-LM runtime, ObjectBox 6 + AppSearch hybrid RRF vault, System-1 decision engine, Prompt Injection Shield, Koog agent harness, 3-tier health bridge, and Firebase AI Logic cloud tier.
• Getting Started & User Manual: Complete guide for installing, debugging over Wi-Fi ADB, sideloading LiteRT-LM models, navigating the 4-card AMOLED deck, invoking watch skills, and developer desktop tools.
• Architecture Research Conversation Log: Detailed chronological decision ledger and deep research synthesis.

---

Architectural Highlights

Layer	Legacy OpenMMPA (Flutter)	WristMind (ai.wristmind.app)
Target Form Factor	Phone Flutter widget tree	Micro-Wrist AMOLED Deck (sw426dp on >2" watch): 80dp touch targets, 20–26sp typography, live /proc/meminfo + PSS RAM HUD pill
Multimodal Embeddings	Unbundled / Cloud API	EmbeddingGemma 2 Text+Vision (440M, 388 MB) & Text (270M) via LiteRT-LM 0.18.0 EmbeddingEngine, sliced via MRL (768d -> 256d + $L_2$ norm, FP16 disabled)
Hybrid Memory Vault	objectbox_flutter_libs 4.0.3 (vector only)	ObjectBox 6.0.0-beta3 HNSW (dimensions = 256, neighborsPerNode = 16, COSINE) + Android 15 Jetpack AppSearch 1.1.0 (PlatformStorage BM25F) fused via Reciprocal Rank Fusion (RRF) + Pre-LLM Grounding Gate
OS Agent Exposure	None	Android 16+ Jetpack AppFunctions 1.0.0-beta01 (@AppFunctionServiceEntryPoint exposing searchMemories, saveMemoryNote, and readWristVitals to system agents)
Local Model Gallery	None	Switchable LiteRT-LM 0.18.0 gallery: Qwen3-0.6B (INT4), FunctionGemma 270M, Gemma 3 270M-IT, SmolVLM-256M-Instruct, SmolVLM2-500M-Instruct with CO-RES / AUTO-SWAP residency
System-1 Router & Shield	Cloud LLM prompt routing	Sub-30 ms Jev-style System-1 Decision Engine (Choice, Score, Noul, confidence gates $P \ge 0.80$ / $0.50..0.80$ / $<0.50$) + LiteRT 2.3.0 CompiledModel Prompt-Injection Shield (Chokepoints A & B)
Agent Harness & FAW Skills	Custom Dart state	JetBrains Koog 1.3.0 (DynamicPromptExecutor, KV-cache History Compression, discoverSkills progressive SKILL.md XML catalog across 7 FAW watch skills)
FAW Health & Sensor Bridge	None	3-Tier Battery-Aware BetterHealthTrackerBridge: Tier 1 TraX V2 ContentProvider (content://com.ailife.betterhealth.healthprovider, zero LED drain), Tier 2 Native Android Sensor.TYPE_HEART_RATE (21) one-shot sampling with immediate unregisterListener, Tier 3 legacy GET_DATA_API + proprietary HSC/CD vendor sensor IDs (31, 43, 65596, 65598, 65599)
Cloud & Speech I/O	STTSFunctions Cloud Functions + TTSEncodings.xlsx	Firebase AI Logic (firebase-bom:35.1.0) + App Check (PlayIntegrity / Debug) + Remote Config + 1 fps Gemini Live camera/voice burst; offline fallback to Android SpeechRecognizer & TextToSpeech

---

Bundled Watch Skills (assets/skills/*/SKILL.md)

1. set-watch-alarm (run_intent): Sets watch alarms (ACTION_SET_ALARM) and countdown timers (ACTION_SET_TIMER).
2. read-wrist-sensors (run_intent): Reads FAW heart rate (BPM), SpO2, steps, calories, distance via TraX V2 ContentProvider + native Sensor.TYPE_HEART_RATE (21) + legacy GET_DATA_API fallback, plus battery % and live PSS RAM.
3. log-vitals-to-vault (run_intent): Samples live or latest FAW vitals and persists a timestamped health note into ObjectBox 256d HNSW + AppSearch BM25F so vitals history is searchable offline.
4. toggle-watch-torch (run_intent): Toggles the FAW camera LED flashlight (CameraManager.setTorchMode) when hardware flash is available.
5. control-watch-system (run_intent): Adjusts watch STREAM_MUSIC volume (UP, DOWN, MUTE, UNMUTE, MAX, %) and opens Android 15 Quick Settings panels (WIFI, BLUETOOTH, DISPLAY, BATTERY, SOUND).
6. schedule-watch-event (run_intent): Schedules watch calendar events via CalendarContract.Events (ACTION_INSERT).
7. quick-js-calc (run_js): Evaluates math and unit conversions inside an isolated, network-disabled WebView sandbox (window.ai_edge_gallery_get_result).

---

Quick Start & Build

# Run unit tests (256d MRL math, FP16 NaN guard, RRF fusion, Prompt Shield, System-1 gates, TraX V2 + Native HR formatting) & build APK
./gradlew testDebugUnitTest assembleDebug

Optional: Push .litertlm / .tflite Weights via ADB (or Download In-App from the MODELS Card)

WristMind boots and runs its seeded Hybrid Vault, System-1 Router, OCR Prompt-Injection Shield, and all 7 Watch Skills immediately out of the box even before model weights are downloaded. To enable neural on-device generation/embeddings offline via adb push:

adb push Qwen3-0.6B.litertlm /sdcard/Download/
adb push embeddinggemma-2-text-vision-440m.litertlm /sdcard/Download/
adb push smolvlm-256m-instruct.litertlm /sdcard/Download/

Attaching Firebase AI Logic (app/google-services.json)

Once app/google-services.json is placed in app/, Gradle automatically activates com.google.gms.google-services, enabling Firebase App Check, Remote Config, and Gemini 2.5 Flash / 1 fps Live Cloud Burst.