The Pipeline That Keeps an AI Tutor From Inventing Math Problems

A week ago, the author shared an update regarding Space Academy, an AI-driven tutoring app designed for their second-grade child. The app focuses on a unique aspect: ensuring that generated content remains accurate, especially for a young learner who might accept errors from the AI. The numbers reflect the app's capabilities: it now covers 41 pages across two textbooks, comprising 199 tasks and 214 check questions. Each task originates from a scanned copy of the physical textbook, subject to strict verification rules.

Verification is paramount, with a vision model transcribing tasks from the scanned pages and a secondary verification comparing the transcription against the original scan. In instances where the model's transcription fails to match the scan, the scan is deemed authoritative. Tasks that involve measurements, such as clock readings or ruler measurements, lack reference answers in the app to prevent potential inaccuracies from propagating through the exercise chain.

The verification process is executed within the app's deployment, not within the AI model itself, ensuring that any attempt at prompt injection would be thwarted due to the CI-based checks. Other observations from the app's usage include:

1. When presented with a question like "Pick four correct inequalities out of six," the app treats it as a set, not a sequence, emphasizing that the order of answers is irrelevant.

2. The AI provides context-specific responses. For instance, if a child incorrectly applies a concept to a situation like "shallow pond / shallow rain," the app would respond with "Rain is never deep, here shallow is about drop size," rather than a generic "incorrect" answer. This approach aims to educate rather than punish.

3. The app also employs creative methods to engage children, such as circular examples where each answer leads to the next card, forming a loop that signals completion upon returning to the starting card. A mistake in this sequence is immediately apparent.

The app also prioritizes practicality by integrating tools that aid learning, such as a physical ruler displayed on-screen for math-related tasks. Additionally, a photo feature allows children to capture and share completed pages with an instant system share option, simplifying the process compared to traditional methods.

For security, SpaceLocker serves as a native guard implemented as a self-contained C# executable. This guard ensures that daily quotas (set by parents) are adhered to, limiting access to other applications like Steam and Roblox, and blocking YouTube via the hosts file. The app also requires a PIN entry for any attempts to terminate the process, and it checks the app's homework status through the Space Academy API when the daily quota is not yet complete.

The author emphasizes that the core technology behind Space Academy is standard and familiar (React, TypeScript, Tailwind, Express, SQLite, and .NET for the guard). However, the real challenge lies in adhering to a disciplined content pipeline, which treats the scanned textbook as an unassailable source of truth, enforces this through build gates, and allows for honest verification where necessary.

This combination of robust content verification and user-friendly features is rare and essential for building a trustworthy educational tool for children.