Self Learning Agentic GPT for Schools & University
What Inspired Me
I noticed that many students use LLMs to copy answers without truly understanding the concepts.
I wanted to build a system that ensures learning actually happens & not just answer generation.
My goal was to create an AI tutor that:
- Refuses to hallucinate
- Detects difficult questions
- Tests understanding through quizzes
- Learns safely from students
What I Learned
I learned that strict Retrieval-Augmented Generation (RAG) is essential to prevent hallucinations.
If context is not available, the system simply says: I don't know
I also learned that active recall improves learning.
So after 3 difficult questions:
$$ \text{if } difficult_count \geq 3 \Rightarrow \text{trigger quiz} $$
This forces students to reflect instead of passively reading answers.
How I Built It
- Flask for backend and session management
- Ollama (Qwen models) for embeddings and LLM responses
- Qdrant as a vector database
- AutoGen agents for:
- Difficulty classification
- Strict RAG answering
- Quiz generation & grading
- Knowledge extraction & verification
- Difficulty classification
When the system doesn't know something, it asks the user to teach it.
Verified facts are embedded and stored:
$$ vector = Embed(fact) $$
This makes the system grow over time.
Challenges I Faced
- Preventing hallucinations completely
- Handling malformed JSON from LLM outputs
- Managing async event loops inside Flask
- Safely verifying user-provided knowledge
Final Outcome
I built a self-learning AI tutor that:
- Encourages deep understanding
- Tests knowledge dynamically
- Admits uncertainty
- Expands its knowledge safely
This project is about transforming AI from a shortcut tool into a true learning partner.