Amber

The Problem Space

Individuals navigating early-stage dementia experience disorientation, face recognition fatigue, and medication confusion. Cloud-reliant AI assistants present unacceptable privacy risks for sensitive domestic environments and fail completely without active Internet access.

The Architecture & System

Engineered an entirely local, always-listening Android companion utilizing on-device Gemma 4 multimodal models. Amber processes camera input, ambient speech (via Vosk), and memory graphs locally without transmitting biometric or personal data off-device.

Native Android application written in Kotlin with Jetpack Compose. Coordinates an on-device quantized Gemma inference pipeline, a low-latency foreground audio capture service, and an encrypted local graph database for facial recognition and routine reminders.

Key Technical Decisions

• Chose on-device Gemma over remote APIs to guarantee zero-leakage patient privacy and strict offline operation.
• Implemented an offline speech-to-text pipeline using Vosk to eliminate cloud transcription dependencies.
• Designed a high-contrast, calm interface minimizing cognitive load for neurodivergent and elderly users.