Saathi

The problem

People with early Alzheimer’s forget mid-task, mistake strangers for relatives, and lose the ability to live safely at home — and existing help either fails the dignity test or sends private home life to the cloud.

How it works

A phone worn on a lanyard that listens, sees, and tracks time, feeding a short-term context log to an on-device Gemma model that decides to observe, remind, or warn — and speaks back in the local language. Three scenarios: task continuity (“what was I doing?”), time-bounded safety (a silent timer when the geyser goes on), and a visitor identity check against enrolled faces. Grounded retrieval only — a hard “I don’t know,” never a fabricated memory.

Key features

• Fully offline, on-device inference — zero telemetry, the family’s life stays home
• Ambient task-continuity recall from mic / camera / time context
• Face-based visitor verification against an enrolled family roster
• Local-language (Hindi) speech in and out
• A LoRA-fine-tuned guardian model with an adversarial hallucination eval set

Architecture

A single Android APK: sensors → Gemma E2B (4-bit) via MediaPipe LLM Inference → short-term context log → decision → Hindi TTS. Faces via ML Kit; state persisted in Room DB.