Blog

June 28, 2026
Predicting danger before it happens: building Vigie, a world-model safety watchdog for robots 🏆 Hackathon Winner

A technical deep-dive into Vigie — the project that won 1st Global Prize at the YC Paris × Unaite Hackathon. An embodiment-agnostic, external safety layer that watches any robot through ordinary cameras, imagines its near futures with a world model, and vetoes dangerous actions before they occur. Hybrid perception (YOLOE + SAM + MediaPipe), a deterministic reflex layer, a Claude-vision generalist judge, and a world-model track built on LeWorldModel (LeCun et al., 2026) for out-of-distribution danger.

August 15, 2025
Memory is all you need

Why episodic memory is the next transformative step for embodied AI agents — and how memory-augmented frameworks let robots situate themselves in a personal history of experience, moving beyond reactive systems toward temporal reasoning.