We’re building an AI Mentor to make personalized
1-to-1 education more accessible, with text tutoring, WebRTC voice chat, OCR and an animated avatar.
Our proposed stack includes Room for offline storage/cache, ML Kit for OCR, a tiny on-device Gemma model for lightweight tasks, Gemini for complex reasoning and cloud voice, and Redis for server-side caching.
The challenge is supporting 3–4 GB RAM phones with intermittent connectivity. Our current local model/runtime uses roughly 700–750 MB RAM, while smaller models offer limited reasoning.
What model, quantization and edge–cloud routing approach would you recommend to reduce cloud-token costs while preserving tutoring quality and device performance? Would love to hear from anyone who has tested similar setups on low-end Android devices.