⚡ Alert Summarizer Agent

An AI-powered agent that automatically clusters, summarizes, and prioritizes operational asset alerts — helping engineering and operations teams focus on what matters most.

Built as part of an application for the AI Systems Internship at Xempla, this project demonstrates how large language models (LLMs) and embeddings can be used in real-world operations & maintenance (O&M) contexts to reduce alert fatigue, surface high-impact issues, and learn from operator feedback.

---

📌 Features

✅ Intelligent Alert Summarization

• Clusters similar alerts using sentence embeddings.
• Summarizes each alert cluster into clear, human-readable insights using Hugging Face's facebook/bart-large-cnn model.

✅ Feedback Loop

• Users can rate summaries as "Helpful" or "Not Helpful" and leave optional comments.
• Feedback is stored and analyzed to monitor and improve summarization quality over time.

✅ Lightweight Streamlit UI

• Upload alert files, view summaries, and submit feedback — all in an intuitive interface.

---

🖼️ Demo

➡️ [https://www.loom.com/share/3dd1fedcd0c34decb970fa6b5a0bb52c?sid=85f65b08-495d-490b-aa4f-34dd6b1eecd7] This short video explains the project, use case, and demo of the prototype.

---

🧠 Use Case Alignment: Why It Matters

This agent is inspired by real problems in O&M environments:

• Alert overload: Engineers are flooded with repetitive or low-impact alerts.
• Slow triage: Identifying what matters takes time and judgment.
• Lack of learning: Systems don't improve based on user feedback.

This agent solves all three, aligning with Xempla's mission of building intelligent, autonomous O&M platforms.

---

📂 Project Structure

alert-summarizer-agent/
│
├── data/
│   ├── sample_alerts.csv       # Sample input alerts
│   └── feedback_log.json       # Stored user feedback
│
├── notebooks/
│   └── alert_summarizer.ipynb  # Jupyter notebook for testing core logic
│
├── src/
│   ├── summarizer.py           # Main summarization logic
│   └── feedback_loop.py        # Feedback storage and stats
│
├── app/
│   └── streamlit_app.py        # UI to interact with summaries and submit feedback
│
├── .env                        # Hugging Face API token (not committed)
├── requirements.txt            # Python dependencies
└── README.md                   # Project overview

---

🚀 How to Run

1. Clone the repo & set up environment

git clone https://github.com/yourusername/alert-summarizer-agent.git
cd alert-summarizer-agent
python -m venv venv
source venv/bin/activate      # Windows: venv\Scripts\activate
pip install -r requirements.txt

2. Add your Hugging Face API token to .env file

HUGGINGFACEHUB_API_TOKEN=your_token_here

3. Run Streamlit app

streamlit run app/streamlit_app.py

🧪 Sample Input Format (sample_alerts.csv)

timestamp,message
2024-06-30 10:30:00,High temperature in AHU-3 zone
2024-06-30 10:31:00,Repeated vibration warning on Pump-2
2024-06-30 10:32:00,Power surge detected in Panel A

🛠 Tech Stack

Area	Tools Used
Language Models	Hugging Face Transformers
Embeddings	sentence-transformers (MiniLM)
Vector Grouping	KMeans Clustering
Frontend	Streamlit
Feedback Storage	JSON

📈 What's Next (Ideas for Iteration)

• Use historical resolution data to suggest automated fixes.
• Rank summaries by severity using external KPIs or sensor data.
• Fine-tune a summarization model on facility maintenance logs.
• Deploy as a Slack bot for real-time feedback collection.

🙌 About the Creator

This project was created as part of my application to the AI Systems Intern role at Xempla. I'm passionate about the intersection of LLMs, real-time operations, and sustainability, and would be excited to help shape the future of autonomous O&M systems.