-
-
Building buddy interface
What Inspired Me Honestly? I signed up for this hackathon without knowing anything about construction. But during research, I watched YouTube videos of Jakarta contractors manually going through architectural drawings, counting materials by hand, then spending hours calling suppliers across the city. It looked painfully tedious - the kind of repetitive work that screams "AI could do this better."
I kept thinking: these people are building actual buildings while stuck with procurement workflows from 1995. That seemed wrong. So I decided to see if I could automate the boring parts.
What I Learned Multi-agent systems are surprisingly powerful - Instead of one massive AI brain, having specialized agents (PDF analysis, material standards, supplier matching) work together creates something more reliable and easier to debug.
Domain expertise matters more than I expected - Learning Indonesian SNI building codes, Jakarta geography, and real construction terminology wasn't optional. Contractors can immediately tell when software was built by someone who actually understands their world.
Google ADK has a learning curve - The Agent Development Kit is powerful but required figuring out specific patterns for agent communication that aren't always obvious from docs.
How I Built It Architecture Coordinator Agent: Orchestrates the workflow PDF Parser Agent: GPT-4o Vision extracts materials from drawings Material Intelligence Agent: Applies Indonesian SNI standards and local terminology Supplier Matching Agent: BigQuery geographic queries find Jakarta suppliers Tech Stack Google ADK for multi-agent orchestration GPT-4o Vision for document analysis BigQuery for supplier database with geographic calculations Streamlit for contractor-friendly UI Cloud Run for deployment Data Integration I researched actual Jakarta construction suppliers, Indonesian SNI building codes, and authentic material terminology. The system uses real supplier coordinates for distance calculations and proper Indonesian phone number formatting.
Challenges I Faced ADK Agent Communication: Spent hours discovering that real ADK agents use different calling patterns than I expected. Direct tool invocation worked better than the coordinator.process() approach I initially tried.
Authentication Hell: Google Cloud permissions took way longer than expected. Needed multiple roles just to run BigQuery geographic queries.
Indonesian Construction Standards: Learning SNI codes, local material names, and Jakarta business practices while building the system. Turns out Indonesian building standards are surprisingly comprehensive and specific.
Time Pressure: Building a working multi-agent system, learning Indonesian construction context, AND creating a decent UI in 10 hours was... ambitious. Had to prioritize core functionality over polish.
Real-world Accuracy: Balancing AI capabilities with authentic Indonesian construction practices. The system needed to feel legitimate to actual contractors, not like a tech demo with fake data.
What Actually Works Building Buddy processes Indonesian construction drawings in ~8 seconds and produces:
Material lists with proper SNI codes and local names Jakarta suppliers ranked by distance from project location Real contact information and pricing estimates Alternative material suggestions for cost optimization The most satisfying part? It genuinely solves a real problem that takes contractors 2-3 days of manual work. Sometimes the best way to learn about an industry is to dive in completely unprepared and see what you can build.## Inspiration
Log in or sign up for Devpost to join the conversation.