Shadow Classroom

Shadow Classroom started from a familiar classroom problem: you rarely know in the moment whether what you just said actually landed. You only see weak signals, a few faces, a couple of questions, and by the time you realize you lost people, you have already spent the time and some students have already checked out. We wanted something that gives teachers a fast feedback loop without replacing the teacher or turning teaching into a chatbot.

So we built a real-time "shadow classroom" that listens to each teacher explaination and evaluates it with a fixed rubric across six learner profiles (like english as second language, anxious, distracted, oppositional, and quiet-but-capable). It estimates learning gain, confusion risk, engagement, and emotional safety, then suggests a small set of genuinely different next moves (diagnose, scaffold, regulate) and shows the tradeoffs for different students.

The part we were most proud of was making the feedback measurable instead of hand-wavy. We generate an influence map by testing small counterfactual edits to the teacher’s sentence, re-scoring, and highlighting the exact spans that drive confusion or harm inclusion. On top of that, we added curriculum coverage using a hybrid RAG knowledge graph built from lecture PDFs and slides, so teachers can see what concepts were actually covered in the conversation versus what the materials intended. Along the way we learned that reliability beats cleverness: strict schemas and deterministic scoring were the difference between a cool demo and something you could actually trust in the middle of a lesson.