We build a local, privacy-preserving language model giving live formative feedback on undergraduate math coursework. A strong open teacher model (e.g. Qwen2.5-Math) is distilled into a compact 3–7B student running on-prem for GDPR-compliant, low-latency use. Students receive structured hints and Socratic type step-gating; teachers get dashboards surfacing recurring misconceptions. Outputs: working model, inference script, guidelines, evaluation report.