SubStrata

A satellite classifier that gets smarter from its mistakes — without ever retraining the model. SubStrata labels satellite imagery (Sentinel-2, Google Dynamic World, EuroSAT) into land cover — trees, shrub, grass, crops, water, built — and paints it as a segmentation map. The twist is what happens when it’s wrong: a second Gemini “Strategist” looks at the miss and writes a plain-English lesson, like “tell trees from shrub by tall shadows and coarse canopy texture,” then files it in a graph organized by a land-cover hierarchy and injects the relevant lessons into the prompt before the next pass. The model’s weights never change — it improves by remembering, not retraining. And because the memory is a hierarchy instead of a flat list, a lesson learned on trees vs shrub automatically transfers to a never-seen grass vs shrub pair (both are vegetation), something a vector database can’t do. Every label can cite the rule behind it, so you can read why it believes what it believes — and an analyst can click a wrong cell, correct it, and watch that correction become a new lesson, human-in-the-loop, weights still frozen. It matters most where you can’t retrain in the field: satellites in orbit, denied-comms defense work, environmental monitoring with a handful of expert corrections instead of a labeling campaign. Land cover is the petri dish; the edge is the point. Built on Gemini 3.5 (Vertex AI), Gemma, Google Earth Engine, a networkx in-memory graph, and a live Next.js dashboard.