Mastery Coding

Sandboxed Python execution on Kubernetes with AI feedback on every student submission — LLM outputs confidence-gated before reaching classrooms.

The Challenge

K–12 students submit untrusted Python code that has to run safely — and the AI feedback on every submission has to be reliable enough for classrooms, not just demos. The feedback layer raised the stakes further: an AI hint that's wrong doesn't just fail to help a student — it actively teaches the wrong thing.

Our Approach

Sandboxed execution on Kubernetes with per-run isolation, and LLM feedback confidence-gated through LangChain so low-confidence output never reaches a student. Every execution gets an isolated Kubernetes sandbox, so no student's code can touch another's session. LLM feedback runs through LangChain with confidence gating — low-confidence responses are suppressed rather than shown.

What We Built

A secure interactive code editor for K–12 students; per-run sandboxed Python execution on Kubernetes; LLM feedback on every submission — code explanation and debugging hints; and confidence gating via LangChain before anything reaches a classroom.

Key Decisions

A sandbox per execution — untrusted student code runs in isolated Kubernetes sandboxes: a runaway loop or malicious snippet can never touch another student's session. Confidence-gated feedback — LLM output is confidence-scored before display; a wrong hint to a 12-year-old is worse than no hint, so low-confidence responses never ship.

The Outcome

Concurrent code executions scaled 20 to 100, with AI feedback on every student submission — gated, classroom-safe, and in production. Concurrent executions scaled five-fold while keeping the isolation guarantee — and the AI feedback shipped classroom-safe instead of dying as a demo. Engagement 2022–2024.