SOLTutor.ai

Inspiration

I'm a dad of two boys — one in elementary school and one in middle school. Over the past year, I started using AI to help them study and prepare for tests, experimenting with different styles of learning and studying techniques. The results were real: my son in middle school made the principal's list and carried a 4.0 GPA all four quarters this year. That experience made me ask a bigger question: what if every student in Virginia had access to this kind of AI-powered learning — not as a chatbot, but as something more immersive? Something that doesn't just quiz them but makes them think, and then coaches them on why their thinking matters? Teachers spend hours grading and guessing which concepts didn't land. Students take a test, get a score, and move on — never really understanding where their reasoning broke down. I wanted to bridge that gap: give students an experience that feels like a game but teaches like a tutor, and give teachers instant, actionable data on exactly what to reteach tomorrow.

What it does

SOLTutor.ai is an AI-powered learning platform aligned to Virginia's Standards of Learning (SOL). Students step into short, branching decision simulations where they don't just pick answers — they make choices and explain their reasoning in their own words. An AI coach analyzes each student's written reasoning in real time, detecting specific misconceptions, evaluating the depth of their thinking, and delivering personalized Socratic coaching feedback — not just "right" or "wrong," but why their reasoning matters historically.

On the teacher side, a live dashboard instantly surfaces which misconceptions are most common across the class, identifies students who chose correctly but for the wrong reasons ("surface-level correct"), and recommends specific reteaching activities — all within minutes of students completing the simulation.

The platform also features:

• Adaptive reading levels that adjust content for below, on, and above grade level
• Voice dictation so younger students can speak their reasoning instead of typing
• AI-generated historical illustrations that bring each decision to life visually
• Reasoning scaffolds — real-time Socratic nudges that appear while students are writing, pushing them to think deeper without giving away the answer

How we built it

The application is built as a single-page React app using TypeScript and Vite for fast development. The UI is styled with Tailwind CSS using a custom dark-mode design system inspired by minimalist aesthetic.

The AI layer is the core of the project. Student reasoning is analyzed through two paths:

1. Live Gemini analysis — student text is sent to the Gemini API with a structured JSON schema that returns misconception tags, evidence keywords, reasoning quality scores, coaching feedback, and confidence bands.
2. Local rule-based fallback — a pattern-matching engine with keyword detection and antonym filtering ensures the app always works, even without an API key. This was critical for reliability. Historical illustrations are generated per-decision using a two-step pipeline: xAI first generates a structured scene specification (setting, characters, action, mood, palette), then an image generation model renders a child-safe watercolor-style illustration.

Challenges we ran into

Misconception detection is harder than right/wrong grading. The biggest challenge was building an AI system that catches why a student is wrong, not just that they're wrong. A student can pick the historically correct answer but reveal a misconception in their reasoning — and that's actually more dangerous than a wrong answer, because it looks like understanding on a traditional test. Getting Gemini to reliably detect these "surface-level correct" responses required careful prompt engineering and structured output schemas. Reading level adaptation. Making the same historical content accessible to a struggling 3rd grader and a gifted 5th grader simultaneously required thoughtful content scaffolding — not just simpler words, but fundamentally different framing and sentence structures. Keeping AI coaching supportive, not punitive. The coaching tone had to walk a fine line: honest enough to challenge misconceptions but encouraging enough that a 9-year-old doesn't shut down. Tuning the Gemini system prompts to hit that balance took many iterations. Image generation safety. Historical content about colonialism, conflict, and slavery requires extreme care when generating illustrations for elementary students. Every image prompt includes child-safety constraints, and the scene specification step acts as a content filter before any image is rendered.

What we learned

The most important thing I learned is that AI is most powerful in education when it analyzes reasoning, not just answers. Multiple-choice tests tell you what a student picked. This system tells you how they think — and that's where real teaching happens. I also learned that the teacher side matters as much as the student side. The most enthusiastic response I've gotten isn't about the student simulation — it's about the teacher dashboard. Teachers are desperate for tools that show them what to reteach before the SOL test, not after.

Built With

• gemini
• pdf.js
• react
• typescript
• vite
• xai