Why Not Just Use ChatGPT?
The case for an AI tutor that actually teaches
Every week there's a new AI wrapper claiming to revolutionize education. Most are pretty UIs around a ChatGPT call. AICadMe is different — not because we use AI, but because we solved the hard problems that raw LLMs can't fix alone.
The Real Problem: LLMs Can't Teach
Ask ChatGPT "explain torque" and you'll get a correct, well-written paragraph. Ask again tomorrow and you'll get a different paragraph. There's no curriculum, no progress, no check for understanding, no adaptivity. Giving answers isn't teaching.
❌ No Curriculum
You have to figure out what to study and in what order. Each session starts from zero.
✅ AICadMe generates a complete syllabus
One prompt → structured subjects, topics, subtopics with a logical teaching order. Built-in, not improvised.
❌ No Teaching Method
LLMs answer questions. They don't have a pedagogical strategy — no Socratic dialogue, no Feynman technique, no progression.
✅ Method-based teaching
Socratic, Feynman, Lecture, Problem-first — each with a crafted system prompt that shapes how the AI tutors. Auto-matched to the subtopic.
❌ No Persistence
Chat history is ephemeral. Close the browser and your learning journey is gone. No understanding tracking across sessions.
✅ Cross-session state
Understanding scores per subtopic persist across sessions. Whiteboard state, notes, and conversation history are all saved. Pick up where you left off.
❌ No Diagrams
You have to explicitly ask for visuals. Most students don't know to ask, and even then the AI describes rather than draws.
✅ Auto-generated diagrams
Every topic opens with a custom diagram rendered inline. No extra prompt needed.
❌ No Progress Tracking
The LLM doesn't know what you know. It can't tell you which topics you've mastered and which need review.
✅ Subtopic-level analytics
Each subtopic has an understanding score (0-100) that updates as you chat. Auto-completes at 80+. Parent dashboard shows everything.
❌ No Structure
Conversations wander. You might spend an hour on one tangent and never cover the actual topic.
✅ Guided teaching plans
Each subtopic has an AI-generated teaching plan that keeps the tutor on track. Key concepts are covered in order. Practice questions are embedded.
What Makes This Hard
Building a real AI tutor isn't about prompts — it's about infrastructure.
Context Management
A single tutoring session can span dozens of messages with long diagrams, teaching instructions, and user profiles. We built a custom context management system with token-aware trimming, summarization, and vector store retrieval — all invisible to the user. Without this, every session blows the token limit after 10 minutes of chat.
Teaching Methodology Injection
You can't just tell an AI to "teach using a specific method" in a system prompt. We spent months crafting instructions that actually make the AI follow a teaching method consistently: asking guiding questions instead of giving answers, breaking problems into steps, and knowing when to switch from explanation to practice. The difference between "sounds like it's using a method" and "actually teaching with that method" is thousands of words of carefully engineered instructions.
Reliable Diagram Generation
Getting an LLM to generate diagrams that actually render properly was a significant engineering challenge. We created a code block convention with JSON metadata and styling guidelines baked into every prompt. The AI doesn't just describe — it generates structured visual content.
Understanding Quantification
How do you measure understanding from a conversation? We built keyword detection that incrementally adjusts scores based on student responses, combined with a completion system that updates topic and subject progress hierarchically. It's not perfect, but it's real — and it works well enough that students feel a genuine sense of progress.
Curriculum Generation at Scale
Generating a complete multi-subject curriculum with topics, subtopics, descriptions, and difficulty levels — all in valid JSON — requires careful prompt engineering, response validation, and error recovery. A single malformed JSON response from the AI can break the entire flow. We built validation layers, retry logic, and a human-readable error reporting system.
What We've Built
- Fully functional AI tutor with SvelteKit frontend and Python/FastAPI backend
- Multi-provider AI support (Gemini, OpenAI, Claude) with automatic fallback
- Custom syllabus generation from natural language prompts
- Goal marketplace for community-curated curricula
- User authentication, parent dashboard, progress tracking
- Interactive simulated environments and quizzes tied to the syllabus
- Custom context management system with vector store retrieval
- Diagram generation integrated into chat
- Persistent whiteboard and notes per subtopic
- Teaching method system with auto-matching to content
- Dockerized production deployment
- Subscription system with usage tracking
Join the Beta Testing Phase
AICadMe is currently in beta and free to use. Sign up now to get early access and help shape the future of AI tutoring.
Join the BetaMarket Timing
The AI-in-education market is projected to reach $112B by 2034. Students are already using ChatGPT — the question isn't whether AI will be used in education, but whether it will be used to learn or to cheat. AICadMe is positioned as the ethical AI tutor that parents and schools can actually feel good about. We turn AI from a shortcut into a teacher.
FAQ
Contact
Interested in investing or partnering? Reach out or use the form below: