StemBuddy

StemBuddy is a production web app where kids run real science investigations — form a hypothesis, design a fair test, collect evidence, defend a conclusion — and file a signed Field Report their parent can read. The hard constraint drove the whole architecture: AI scaffolds every step but never ghostwrites, auto-grades, or offers a one-tap accept. That rule is enforced in code, not in a style guide.

The enforcement is concrete. Every AI “stronger version” of a kid’s writing renders through one shared component that opens on a word-level diff with a written rationale, then offers three balanced choices — use this, keep mine, combine — with no lone accept button. The kid is authoring, not approving, and the evaluation itself is logged as the skill being practised.

Behind it sits a skill-based coach orchestrator: ~16 routed action handlers, composable context skills, stage isolation so hypothesis context can never leak into procedure flows, and a per-user fact-extraction cap so the coach gets more personal over time without growing unbounded.

Policy compiles here too. COPPA and privacy commitments are executable gates, not review discipline: a build fails if any third-party tracker appears in the kid-facing bundle, and secret, content-log, accessibility and design-system validators run in the same gate.