🏗️ PCL Bot: AI-Powered Assistant for Engineers in the AEC Industry

📘 Overview

PCL Bot is a domain-specific AI solution built to streamline how engineers in the Architecture, Engineering, and Construction (AEC) industry access and interact with technical documentation. Focused on standards like ASME B31.3, this digital assistant answers complex engineering queries with precise, referenced, and well-structured responses.

---

❓ Problem Statement

Engineers often spend excessive time sifting through bulky binders and PDFs to locate relevant clauses in technical standards. This:

• Delays decision-making on site
• Increases risk of human error
• Impacts efficiency and safety

---

💡 Our Solution

A 24/7 intelligent assistant that enables:

• 📥 Natural language input from engineers
• 📌 Accurate, summarized responses
• 🔗 Direct references to standard clauses
• ⚡ Rapid access to critical information

---

🧪 Technical Approach

🔹 Base Model

• Mistral-7B OpenOrca — a powerful open-source LLM chosen for its strong performance on QA and summarization tasks.

🔹 Enhancements

1. Retrieval Augmented Generation (RAG)
   • Embeds documentation into a vector database
   • Uses similarity search to inject relevant context
2. Fine-Tuning with LoRA (PEFT)
   • Fine-tuned on ASME B31.3 to improve domain relevance
   • Customized model responses to match the style and structure expected by engineers

🔹 Final Architecture: RAFT

• RAG + Fine-Tuned Model = RAFT
• Improved factual accuracy, precision, recall, and style
• Reduced hallucinations and irrelevant information

---

📊 Evaluation

• ROUGE: Evaluates text overlap with reference answers
• Cosine Similarity: Measures semantic closeness
• F-Score: Combines both for a balanced metric

Weighted F = w1*ROUGE + w2*CosSim
Example Weights: [0.15, 0.15, 0.15, 0.15, 0.15, 0.25]

---

⚙️ Limitations

• RAG struggles with tables unless preprocessed into readable text
• PEFT changes style more than raw accuracy
• No real-time integration with site-specific data yet

---

🚀 Next Steps

• Improve RAG table handling
• Deploy the app for domain expert testing
• Use feedback for another iteration of fine-tuning

---

👨‍💻 Demo

👉 Click here for the Demo

---

👥 Team Members

• Advik Mehta
• Anant Bhide
• Falak Sethi
• Hanzhe Ye
• Shreyank Hebbar

Domain Expert: Brian Gue (PCL Industrial, Data Science)

---

📜 License

This project is for academic purposes and may be adapted or extended with proper attribution.