Lazaro Martull

LeafTutor – AI-Powered Programming Tutor

LeafTutor is an AI-driven tutoring system designed to provide structured, step-by-step programming guidance to post-secondary students.

This system was evaluated through real programming assignments and published as:

🔬 LeafTutor: An AI Agent for Programming Assignment Tutoring
arXiv:2601.02375
https://arxiv.org/abs/2601.02375

---

🧩 My Role

I contributed to:

• System architecture design
• Backend implementation (Flask API)
• LLM prompt structuring for step-by-step reasoning
• Database schema design (PostgreSQL)
• Evaluation and testing workflow
• Documentation and research contribution

---

🎯 Problem

High enrollment in STEM programs has created increasing demand for scalable programming tutoring support.

Traditional tutoring:

• Does not scale well
• Is resource-constrained
• Varies in instructional quality

The goal was to design an AI system capable of delivering clear, structured, step-by-step programming explanations comparable to human tutors.

---

🧠 System Architecture

LeafTutor integrates:

• Frontend → Student query submission
• Backend (Flask) → API handling and session management
• Database (PostgreSQL) → Storing assignments and interactions
• LLM Layer → Structured prompt design for step-by-step reasoning
• Evaluation Pipeline → Comparing outputs against expected tutoring quality

Core Flow

1. Student submits programming question
2. Backend formats structured prompt
3. LLM generates step-by-step instructional response
4. Output is returned with organized explanation
5. Evaluation framework measures instructional clarity and completeness

---

🛠 Tools & Technologies

• Python
• Flask
• PostgreSQL
• LLM API (Gemini / GPT-style models)
• Prompt engineering
• Research-based evaluation framework

---

🧪 Prompt Engineering Strategy

To simulate tutoring behavior, prompts were structured to:

• Break problems into logical steps
• Explain reasoning before providing solutions
• Avoid directly dumping final answers
• Encourage guided learning

This structured prompting significantly improved clarity and instructional usefulness.

---

📊 Evaluation & Results

LeafTutor was evaluated using real programming assignments.

Key findings:

• Delivered structured step-by-step explanations
• Maintained instructional clarity
• Demonstrated guidance quality comparable to human tutors
• Reduced ambiguity in programming explanations

This validated the system’s potential for scalable STEM tutoring.

---

🔍 Research Contribution

This project contributes to:

• Applied AI in education
• LLM tutoring systems
• Instructional design via prompt engineering
• Scalable academic support solutions

The work was published on arXiv and categorized under:

• Artificial Intelligence (cs.AI)
• Software Engineering (cs.SE)
• Computers and Society (cs.CY)

---

🚀 Future Improvements

Planned enhancements include:

• Fine-tuned LLM for tutoring-specific behavior
• Rubric-based evaluation scoring
• Multi-turn conversation tracking
• Personalized learning profiles
• Instructor dashboard analytics

---

📁 Folder Structure

projects/
└── leaftutor/
    ├── index.md              # Project write-up (this page)
    └── leaf_tutor_paper.pdf  # Research paper (PDF)

---

🎯 Summary

LeafTutor demonstrates how large language models can be engineered to deliver structured, pedagogically sound programming guidance.

By combining system design, backend engineering, prompt structuring, and real-world evaluation, this project bridges academic research and applied AI system development.