Every PDF you upload is hashed. Every answer we generate is stored. When the next student asks the same thing, we serve the cached result — no GPU spin-up, no datacenter cooling water, no extra carbon.
The moment a file lands, we compute a SHA-256 of its contents. If that exact hash already exists in our knowledge graph, we skip parsing entirely and link to the prior extraction — saving the whole OCR + topic-extraction pipeline.
Each answer is keyed by user, normalized question text, and mark weight. When the cache key matches, we return the stored answer in milliseconds instead of running the model again.
We meter avoided inference the way utilities meter electricity. 25 mL water and 0.3 Wh per call (industry-published frontier-model figures). Here's what the platform has kept off the grid so far.
Every new student inherits the entire prior knowledge graph. Each upload, each question, each cached answer makes the next request slightly cheaper to serve — for users, and for the grid.