Project Description Project Title: S.T.R.U.C.T. Structural Testing, Reasoning & Unified Computational Tool.
Inspiration Structural engineering often involves cross-referencing complex diagrams with dry technical solvers. We wanted to build a "Jarvis" for civil engineers—an agent that can "see" a loading diagram, "hear" an engineer's verbal constraints, and "speak" back technical insights in real-time, bridging the gap between human intuition and FEM (Finite Element Method) precision.
What it does S.T.R.U.C.T is a multimodal engineering copilot that:
Sees: Analyzes structural diagrams (Shear Force, Bending Moment, or Beam Layouts) using Vision-Language Models. Hears: Transcribes verbal engineering queries in real-time using the Gemini Multimodal Live API. Speaks: Verbally explains stress peaks, safety factors, and material limits using high-fidelity TTS fallback. Simulates: Automatically triggers a custom FEM core to compute SFD, BMD, and Deflection plots instantly. How we built it Core: Python-based FEM solver engine (NumPy, SciPy). Brain: Gemini 2.0 Flash for real-time multimodal reasoning. Frontend: React-based dashboard with Framer Motion and Lucide icons for a premium engineering aesthetic. Backend: FastAPI monolith serving both the API and the static frontend. Deployment: containerized on Google Cloud Run with automated CI/CD for rapid iteration. Challenges we ran into WebSocket Stability: Handling the real-time handshake of the Multimodal Live API required careful state management and robust error fallback to HTTP + Web Speech API to ensure the "Voice" never breaks during a live demo. Parameter Normalization: Converting natural language ("a 2-meter stick with 5 tons on the end") into precise SI units for the FEM solver required a fine-tuned grounding loop. Accomplishments that we're proud of Seamless Fallback: We implemented a "Safety Net" that silently shifts between Gemini Live and Browser Voice if the connection fluctuates, ensuring a 100% stable user experience. Premium UX: The dashboard feels like a high-tech telemetry station, making complex structural data accessible and beautiful. What we learned The power of "Model-in-the-loop" for technical grounding. How to manage high-throughput PCM audio chunks for low-latency voice interaction.
Built With
- ai
- cloud-run
- docker
- fastapi
- framer-motion
- gemini-2.0-flash
- google-cloud
- lucide-react
- matplotlib
- multimodal-live-api
- numpy
- python
- react.js
- scipy
- vertex
- web-speech-api
- websockets
Log in or sign up for Devpost to join the conversation.