Agentic AI Chat Analyzer

An AI-powered platform for analyzing agent chat transcripts about Washington Post articles using lightweight LLMs, visual EDA, and an interactive frontend.

Features

• Modular Data Pipeline – Clean ingestion, cleaning, and transformation
• EDA + Profiling – Dataset summaries, word clouds, sentiment plots
• LLM Summarizer – Uses Flan-T5-small
• Sentiment Classification – CardiffNLP RoBERTa Sentiment
• FastAPI Backend – REST endpoints with Pydantic validation
• Streamlit Frontend – Interactive UI for real-time transcript analysis
• Model Caching – Offline-ready with locally saved models
• Evaluation Pipeline – Accuracy and BLEU score metrics

Project Structure

Agentic-AI-Chat-Analyzer/
├── src/
│   ├── api/              # API endpoints
│   ├── services/         # Business logic (models, metrics, summarizer)
│   ├── utils/            # Data processing utilities
│   ├── core/             # Core modules (logger, main)
│   └── schemas/          # Pydantic models
├── scripts/              # Utility scripts
├── notebooks/           # EDA notebooks
├── streamlit_app/       # Streamlit UI
├── data/                # Dataset files
├── models/              # Saved ML models
├── main.py              # FastAPI entrypoint
└── requirements.txt     # Dependencies

System Architecture

User Input (Streamlit UI)
    ↓
POST Transcript to FastAPI API
    ↓
    ├── LLM Summarizer (Flan-T5)
    └── Sentiment Classifier (RoBERTa)
    ↓
Summary + Sentiment Output
    ↓
Visualization (WordClouds + Metrics)

Installation

1. Clone the repository
2. Create a virtual environment:

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

3. Install dependencies:

   pip install -r requirements.txt

4. Download models:

   python scripts/download_models.py

Usage

Run FastAPI Backend

uvicorn src.main:app --reload

Access Swagger Docs: http://localhost:8000/docs

Run Streamlit Frontend

streamlit run streamlit_app/app_ui.py

API Endpoints

Method	Endpoint	Description
GET	/summary	Returns dataset stats
POST	/transform	Preprocess and clean new chat input
POST	/insights	Summarizes, classifies sentiment, returns insights

Tech Stack

• Backend: FastAPI, Uvicorn, Pydantic
• Frontend: Streamlit, Matplotlib, WordCloud
• NLP Models: HuggingFace Transformers (Flan-T5, RoBERTa)
• Analysis: Pandas, Seaborn, ydata-profiling
• LLM Hosting: Locally saved models in /models/

License

MIT License