🤖 Agent Reasoning Evaluator

🧠 Overview

Agent Reasoning Evaluator is a lightweight web app built with Streamlit that analyzes and annotates reasoning trajectories of AI agents.
It simulates how intelligent agents reason, make decisions, and sometimes make logical or factual errors — allowing users to evaluate and visualize their performance.

The project was inspired by agentic AI concepts and practical insights from workshops on NVIDIA NeMo Toolkit and AgentToolkit.

---

🚀 Live Demo

👉 https://agent-reasoning-evaluator.streamlit.app/

---

🧩 Features

• ✅ Upload agent reasoning data (agent_data.json)
• 🔍 Automatically evaluate outputs for factual or logical errors
• 🧾 Generate structured annotations (annotations.json)
• 📊 View accuracy metrics and error distributions
• 💾 Download annotated results
• 🌐 Clean and intuitive Streamlit interface

---

🧰 Tech Stack

• Python 3.11+
• Streamlit
• Pandas
• JSON

---

📁 Project Structure

agent-reasoning-evaluator/ │ ├── app.py # Streamlit dashboard ├── evaluate_agent.py # Command-line evaluation script ├── agent_data.json # Sample reasoning trajectories ├── annotations.json # Generated annotations ├── summary.csv # Summary results └── README.md

⚙️ How to Run Locally

1. Clone the repository

   git clone https://github.com/<your-username>/agent-reasoning-evaluator.git
   cd agent-reasoning-evaluator

2. Install dependencies pip install streamlit pandas
3. Run the Streamlit app streamlit run app.py
4. Open your browser at locahost

💡 Future Enhancements

Add detailed reasoning step visualization

Integrate with real-world agent reasoning traces

Enable multi-file uploads and bulk evaluation

Add charts for reasoning accuracy trends

🧑‍💻 Author

Chinmay Kadu Web Developer | AI Enthusiast | Agentic Systems Learner 🌐 https://agent-reasoning-evaluator.streamlit.app/ 📧 chinmayrkadu@gmail.com