Micro-Display Ambient Intelligence & Hybrid Agentic Memory (Tiny Phone & Full Android Watch • 4GB RAM)
Kotlin-native agent harness parsing SKILL.md catalogs into XML prompts, managing multi-turn state machines off the main thread.
7 outcome-oriented watch skills (Alarm, Timer, Flashlight, Panels, Vitals) invoked via ML Kit AgentRunner.
Reuses resident EmbeddingGemma text-head with route centroids for sub-30ms confidence gating without allocating extra RAM.
Client-side Firebase AI Logic connection using BoM 35.1. Zero backend cold-starts and direct Gemini 2.5/3 Flash execution.
Escalates to cloud Gemini when local confidence falls below 50% or complex multimodal synthesis is required.
Sanitizes prompt injections, enforces delimiter boundaries, and blocks token exfiltration before cloud dispatch.
Slices 768d vectors to 256d + L2 unit norm, saving 66% vector RAM while maintaining retrieval precision.
Hardware-accelerated lexical search fused with HNSW vectors via Reciprocal Rank Fusion: Score = 1/(k + rank).
80dp OLED deck with headless @Composable AppState Transform HUD & StateStore syncing UI with background AppFunctions.
Voice STT / Camera OCR
< 30ms • 0MB Added RAM
P≥0.80 Local | P<0.50 Cloud
ML Kit AppFunctions OR Firebase
AppState Transform Pill + TTS
Convergence of ML Kit GenAI, JetBrains Koog, Firebase AI Logic (no Cloud Run), Jetpack AppState, AppFunctions, AppSearch, and ObjectBox 6 HNSW.
Solves 2-inch screen typing friction. Instant sub-30ms offline responses, zero cloud latency on local actions, zero battery drain from PPG health sensors.
Agentic memory designed for 4GB RAM tiny phones. 256d Matryoshka truncation fits full vector HNSW indexing inside a 1.2 GB PSS budget.