ArchiShield

Inspiration

Architecture and construction projects generate enormous volumes of technical data — BIM models, CAD drawings, compliance codes, cost sheets — but this information is scattered across specialized tools that only technical experts can interpret. Architects, builders, clients, and municipal reviewers often can't access the same insights in a format that makes sense to them. We built ArchiShield to bridge that gap: a single platform that turns raw architectural data into clear, role-adaptive insights for everyone involved in a project.

What it does

ArchiShield is an AI-powered analysis platform for architectural and BIM data. It includes:

• Compliance analysis — automatically checks building data against regulatory requirements
• Cost estimation — surfaces budget insights from project data
• Quality & risk scoring — flags design issues before they become costly problems
• Role-adaptive dashboards — five distinct UI themes so a technical engineer, a business stakeholder, or a non-technical client each see the view that matters to them
• Natural language querying — an AI assistant (powered by Groq's LLaMA 3.3 70B) that lets users ask questions about their project data in plain English
• Voice input — for hands-free interaction on-site

How we built it

• Frontend: Streamlit multipage app with 13 analysis modules and a Technical/Business/Full view toggle
• Backend: A decoupled FastAPI service (ported from the original Streamlit app) using BackgroundTasks for async job processing and an in-memory job store
• Data & Auth: Supabase for storage and JWT-based authentication
• AI layer: Groq LLaMA 3.3 70B for the natural language query assistant, with prompt/token optimization reducing usage by ~92% (down to ~1,800 tokens per query)
• Deployment: Streamlit Community Cloud

Challenges we ran into

• Decoupling a monolithic Streamlit app into a proper async FastAPI backend without breaking the 13 interdependent analysis modules
• Reducing LLM token usage for the NLQ assistant to keep response times and costs manageable
• A subtle bug in our scoring module (Model_Score.py) where a percentage-based metric could exceed 100% under certain input conditions — caught and fixed during backend migration
• Designing role-adaptive dashboards that stay genuinely useful for both deeply technical and completely non-technical users, rather than just hiding fields

Accomplishments that we're proud of

• Built and deployed a full 13-module analysis platform in a short hackathon timeframe
• Cut LLM token consumption by over 90% through careful prompt engineering
• Designed a UI that adapts meaningfully across five different user roles

What we learned

• How to structure a FastAPI backend around long-running analysis jobs using background tasks
• Practical techniques for reducing LLM API costs without sacrificing answer quality
• The value of catching edge-case bugs (like our percentage overflow) early, before they reach a business-facing dashboard

What's next for ArchiShield

• Porting the remaining analysis modules from the original Streamlit app to the FastAPI backend
• Expanding compliance checks to cover additional regional building codes
• Adding persistent job storage (replacing the in-memory store) for production reliability

Built With

• ai
• architecture
• background-tasks
• bim
• cad
• dashboard
• fastapi
• groq
• jwt
• llama-3.3-70b
• natural-language-processing
• python
• rag
• streamlit
• supabase