Dev Team Builds AI Tutor That Questions Students Instead of Writing for Them

A development team built a Socratic debate agent designed to challenge students' arguments rather than improve their writing, deliberately avoiding the ghostwriting pitfall common in AI tutoring tools. The system uses deterministic Python functions for validation, persona selection, and fallacy clustering, reserving LLM calls only for tasks like generating Socratic questions and qualitative scoring. A security lesson emerged when Google's gmail.compose scope unexpectedly allowed sending emails, prompting the team to remove all send functionality and enforce the restriction via AST-based CI tests. To combat OCR hallucinations in handwritten essay uploads, the team runs Gemini Vision twice on each image and uses string-similarity diffing to override the model's self-reported confidence when outputs diverge. The project highlights a hybrid engineering approach where deterministic logic handles auditable decisions and LLMs are confined to tasks that rule-based code genuinely cannot perform.