Intelligent Learning Analytics & Agentic AI Study Coach

Unified Project: Milestone 1 + Milestone 2

---

Quick Navigation

• Milestone 1 — ML-Based Student Performance Prediction
• Milestone 2 — Agentic AI Study Coach
• System Architecture
• Tech Stack
• Getting Started
• Project Structure
• Roadmap

---

Milestone 1 — ML-Based Student Performance Prediction

Overview

Milestone 1 implements a machine learning pipeline that predicts student academic performance using demographic, academic, and behavioural data.

The system provides three outputs:

• Exam score prediction (regression)
• Pass/Fail classification
• Learner segmentation using clustering

A Streamlit dashboard is used to interactively visualize predictions.

Live Demo:
https://predictive-learning-analytics-ml.streamlit.app/

---

Models & Performance

Model	Task	Performance
Linear Regression	Predict ExamScore	R² = 0.9397
Logistic Regression	Pass/Fail	Accuracy = 91.76%
K-Means Clustering	Learner Segmentation	Silhouette = 0.2112

---

Key Design Decisions

• WritingScore used as proxy target (no leakage risk)
• Median-based Pass/Fail threshold (69.0)
• Manual ordinal encoding for education-related features
• Midpoint encoding for study hours
• class_weight='balanced' instead of SMOTE
• k=3 clustering aligned with interpretability requirement
• Train-only scaling to avoid data leakage

---

Dataset

• 30,640 student records
• 14 original features
• 11 engineered features

Source: Kaggle Students Exam Scores Extended Dataset

---

Preprocessing Pipeline

1. Data cleaning and normalization
2. Missing value imputation
3. Ordinal + one-hot encoding
4. Outlier handling (IQR method)
5. Feature scaling (StandardScaler)
6. Train-test split (80/20)
7. Model training and evaluation

---

Milestone 2 — Agentic AI Study Coach

Overview

Milestone 2 transforms the ML system into a conversational agentic AI tutor.

Instead of static inputs, students interact through natural language, and the system dynamically decides:

• What analysis to run
• What knowledge to retrieve
• Whether to generate a study plan or quiz
• How to respond using ML + LLM reasoning

This is implemented using LangGraph-based multi-agent orchestration.

---

Key Features

1. Conversational AI Interface

Students interact via chat instead of form inputs.

2. LangGraph Agent System

A structured graph of specialized nodes:

• Analyser Node (ML inference)
• Retriever Node (RAG system)
• Planner Node (study plan generation)
• Quizzer Node (MCQ system)
• End Node (response finalization)

---

3. Machine Learning Integration

• Extracts student data from natural language
• Runs all Milestone 1 models dynamically
• Predicts:
  • Exam score
  • Pass/Fail status
  • Learner category

---

4. Retrieval-Augmented Generation (RAG)

• FAISS vector database
• Sentence-transformer embeddings (MiniLM)
• Academic knowledge base
• Optional Tavily web search fallback

---

5. Personalized Study Plans

• 7-day structured study plan
• Generated from retrieved knowledge
• Adapted to learner category:
  • At-Risk → fundamentals + revision
  • Average → balanced learning
  • High Performer → advanced + challenge tasks

---

6. AI Quiz System

• 5 MCQs per session
• Auto-generated from retrieved content
• Automatic grading + explanations
• Performance-based feedback

---

7. Persistent Memory

• PostgreSQL (Neon)
• Stores:
  • chat history
  • agent state
  • session data
• Allows conversation resumption

---

8. Academic Guardrails

• Blocks non-academic queries
• Prevents cheating requests
• Ensures safe tutoring behavior

---

System Architecture

User (Streamlit UI)
|
v
LangGraph Agent (Master Node)
|
v
+----------------------------------+
| Specialist Nodes                 |
| - Analyser (ML Models)           |
| - Retriever (RAG System)         |
| - Planner (Study Plan Generator) |
| - Quizzer (MCQ System)           |
| - End Node                       |
+----------------------------------+
|
v
PostgreSQL (Persistent Memory)

---

Tech Stack

Layer	Technology
ML Models	Scikit-Learn
Agent Framework	LangGraph
LLM	Groq (Llama 3.3 70B)
Embeddings	all-MiniLM-L6-v2
Vector DB	FAISS
Backend	Python
UI	Streamlit
Database	PostgreSQL (Neon)
Deployment	Streamlit Cloud
Web Search	Tavily API

---

Getting Started

Clone Repository

git clone https://github.com/sathvik89/Predictive-Learning-Analytics_ML.git
cd Predictive-Learning-Analytics_ML

---

Create Environment

python -m venv venv
source venv/bin/activate   # Mac/Linux
venv\Scripts\activate      # Windows

---

Install Dependencies

pip install -r requirements.txt

---

Run Application

streamlit run app.py

---

Project Structure

predictive-learning-analytics
├── __pycache__
│   ├── app.cpython-313.pyc
│   └── styles.cpython-313.pyc
├── agent
│   ├── __init__.py
│   ├── __pycache__
│   ├── chat_history.py
│   ├── formatting.py
│   ├── graph.py
│   ├── guardrails.py
│   ├── ml_pipeline.py
│   ├── nodes.py
│   ├── rag.py
│   ├── session_context.py
│   └── state.py
├── app_errors.log
├── app.py
├── chat_history.db
├── Data
│   ├── processed
│   └── raw
├── knowledge
│   ├── academic_coaching.md
│   ├── algebra_geometry_trig.md
│   ├── math_foundations.md
│   ├── performance_intervention.md
│   ├── reading_comprehension.md
│   ├── README.md
│   ├── statistics_probability.md
│   ├── study_skills.md
│   └── writing_skills.md
├── models
│   ├── kmeans_model.pkl
│   ├── linear_model.pkl
│   ├── logistic_model.pkl
│   ├── scaler_clf.pkl
│   ├── scaler_cluster.pkl
│   └── scaler_reg.pkl
├── modules
│   ├── __pycache__
│   ├── components.py
│   ├── home.py
│   ├── icons.py
│   ├── model_loader.py
│   ├── performance.py
│   ├── predict.py
│   ├── sidebar.py
│   └── styling.py
├── notebooks
│   ├── AgenticAI_Practice_Roughbook.ipynb.ipynb
│   ├── Cleaned__Notebook.ipynb
│   └── GenAi_Project_Predictive_learning.ipynb
├── README.md
├── Report
│   └── GenAi_Final_Report_v2.pdf
├── requirements.txt
├── styles.py
└── venv
    ├── bin
    ├── etc
    ├── include
    ├── lib
    ├── pyvenv.cfg
    └── share

---

Roadmap

Completed

• ML-based student prediction system
• Streamlit dashboard
• LangGraph agent system
• RAG pipeline
• Quiz system
• Persistent memory

---

Future Work

• User authentication system
• Step-by-step adaptive quiz flow
• Larger academic knowledge base
• Student analytics dashboard
• Mobile UI optimization
• Personalized long-term learning tracking

---

Project Summary

This project demonstrates the evolution from a traditional machine learning pipeline into a fully agentic AI tutoring system. It integrates predictive modeling, retrieval-augmented generation, and multi-agent orchestration to create a system that not only analyzes student performance but actively supports learning through conversation, planning, and assessment.

---

Built for Gen AI Course — Milestone 1 & 2 Sathvik Koriginja | Anushka Tyagi | Apoorva Choudhary