LLM-Based Course Planner

Personalized academic course planning system powered by a locally-hosted LLM — takes a student's background and target course, then generates a structured step-by-step learning path using four progressively sophisticated prompting strategies.

Live Project Report

Full results, prompt examples, and technique comparison: https://sadhanageddam27.github.io/llm-course-planner/

---

Project Structure

llm-course-planner/
└── LLM.ipynb    # All 4 parts in one notebook

---

What This Project Covers

System Pipeline

Student Input → Prompt Builder → LLM API → JSON Parser → Learning Path

Takes: department, degree level, completed courses, target course. Returns: ordered prerequisite steps mapped to real course catalog entries.

Part 1 — Zero-Shot JSON Prompting

Single prompt instructing the LLM to return only a valid JSON object with prerequisites and a suggested learning path. No examples provided.

Sample output for student targeting Deep Learning:

{
  "recommended_prerequisites": [
    { "course_number": "CSCI 356", "course_name": "Data Structures" },
    { "course_number": "CSCI 343", "course_name": "Data Science" }
  ],
  "suggested_learning_path": [
    { "step": 1, "course_number": "CSCI 632", "course_name": "Machine Learning" },
    { "step": 2, "course_number": "CSCI 492", "course_name": "Deep Learning" }
  ]
}

Part 2 — Retry Logic and JSON Repair

LLMs occasionally return malformed JSON with comments or missing quotes. Added automatic retry — strips comments, sanitizes response, re-queries up to 3 times before graceful fallback. Handles hallucinations reliably.

Part 3a — Chain-of-Thought Prompting

Explicit multi-step reasoning prompt:

1. Identify foundational concepts required for the target course
2. Compare with student background and identify gaps
3. Recommend only courses from the available catalog

Forces the model to justify each recommendation before committing — significantly reduces hallucinated course suggestions.

Part 3b — Multi-Step Decomposition

Breaks the single task into 3 independent LLM calls, each feeding the next:

• Step 1 — Required knowledge topics for the target course
• Step 2 — Gap analysis against student completed courses
• Step 3 — Course recommendations mapped to identified gaps

Each step is independently verifiable and debuggable — the most reliable and explainable approach across all four parts.

---

Technique Comparison

Part	Technique	Reliability	Key Feature
1	Zero-shot JSON	Moderate	Simple, fast
2	Retry + JSON repair	High	Handles malformed output
3a	Chain-of-Thought	High	Explicit reasoning
3b	Multi-step decomposition	Highest	Each step verifiable

---

Setup and Usage

git clone https://github.com/sadhanageddam27/llm-course-planner.git
cd llm-course-planner

# Set your LLM API credentials as environment variables
export LLM_API_KEY="your-api-key"
export LLM_HOST="your-host"
export LLM_PORT="50001"

pip install requests jupyter
jupyter notebook LLM.ipynb

The notebook uses a locally-hosted LLM endpoint. Swap the base URL to point to any OpenAI-compatible API (OpenAI, Ollama, etc.).

---

Tech Stack

Python · Jupyter · requests · JSON · Prompt Engineering · LLM API

Topics

llm prompt-engineering nlp python chain-of-thought json course-planning machine-learning