
=== The problem we couldn't stop thinking about === We grew up in Japan. Typhoons, earthquakes, tsunamis. The alerts saved lives — but the message was always the same: a hundred thousand different people receiving identical words. A warning. Not a plan. People aren't generic. Same typhoon, three different locations, three different bodies, three different instructions. An 82-year-old diabetic in a wood-frame house, a parent driving south with two kids, a wheelchair user on the ground floor — each needs a plan made for them alone. Sonae adds that layer. === What it is === Every citizen gets a personal AI agent. Each agent reasons through their situation — building, family, mobility, medication, location — and generates a plan written for them alone. Citizens can send photos: flooded streets, smoke under a door — and Aki's response shifts based on what it sees. It runs on a single AMD Instinct MI300X (192GB HBM3) at the city edge. 8,000 autonomous agents in parallel, all powered by Qwen3.5-122B-A10B-FP8 — a natively multimodal model handling text and image in one forward pass. Everything runs locally. No cloud. It keeps working when networks fail. === Why this was hard === On Day 1 we set checkpoints: if 100 agents weren't running by noon, fall back to a 30B model. They ran. By evening, 8,000 were thinking in parallel. That moment was the project. A disaster app cannot feel like a disaster — warm off-white, deliberate motion, short specific sentences. When you're scared, you need someone who already knows you. === What's different and next === Cloud platforms personalize at the neighborhood level. Sonae personalizes per individual, fully offline. Networks fail in disasters, and someone with insulin running low cannot wait for the cloud. Phase 2: physics-based hazard models, scaled to million-citizen cities. === Who built this === Two Japanese builders, Vancouver, 80 hours. Beyond alerts. Personal action. Built in Vancouver. For Japan.
10 May 2026