Building Scorpio: Verifiable Physics Tutoring

Physics is not just a collection of formulas; it is a framework for understanding the universe. However, for many students, that framework is obscured by “answer-seeking” behavior. When I set out to build Scorpio, my goal wasn’t to build another “AI Homework Helper.” I wanted to build a Verifiable Physics Tutoring LMS that forces students to think, while ensuring every interaction is grounded in scientific truth.

The Problem: The “Hallucination” Gap in Education

Generic LLMs are dangerous in a physics context. They prioritize plausibility over correctness, often “hallucinating” derivations or skipping crucial unit conversions. In a learning environment, a wrong step is more than just an error—it’s a pedagogical failure.

To solve this, I moved away from simple prompting and engineered a 4-Layer Constraint Architecture.

Engineering for Pedagogical Integrity

The core of Scorpio is a system of “Hard Constraints” that the AI must satisfy before any output reaches the student.

🛡️ The Layered Defense

1. Domain Constraint: Filters out non-physics queries to keep the student focused.
2. Pedagogical (Socratic) Constraint: This is the most critical layer. It programmatically prevents the AI from providing direct answers. Instead, it must identify the student’s specific misconception and offer a targeted hint.
3. Notation & Unit Constraint: Mandates strict adherence to SI units and standard LaTeX formatting for all mathematical expressions.
4. Composite Logic: A final verification pass that ensures the response is consistent with the current problem state stored in our Firestore database.

Technical Implementation: Real-Time & Physics-First

The frontend isn’t just a shell—it’s an interactive laboratory.

• Real-Time Sync: Using Firebase’s snapshot listeners, student progress and AI “thought chains” are synced across sessions with zero perceived latency.
• Mathematical Precision: We integrated a custom KaTeX rendering engine that handles complex multi-line derivations. I also built a Visual Math Builder to help students construct equations without needing to learn raw LaTeX.
• Space-Themed UX: To reduce “physics anxiety,” I designed a high-performance “glassmorphic” interface. It uses Framer Motion for physics-based UI interactions, mirroring the concepts being taught.

Impact & The Path to Davidson Fellows

Scorpio was submitted to the 2026 Davidson Fellows Scholarship. In initial pilot groups at Sage Ridge, we saw a marked shift: students stopped asking “What is the answer?” and started asking “Why does this force vector change?”

The system tracks Rule Adherence % and Response Quality metrics, allowing us to verify that our AI constraints are actually working.

Reflections on AI Engineering

Building Scorpio taught me that the future of AI in education isn’t about more intelligence—it’s about governed intelligence. By setting strict boundaries, we actually create more freedom for the student to explore, fail safely, and eventually, achieve a breakthrough in understanding.

Related Project:

Scorpio — AI Physics Tutoring LMS
Explore the Architecture →