What Inspired Me
It all started two months ago at 3 AM. I am a 19-year-old sophomore studying Mining Engineering in Chile, and I struggle with insomnia. I spent my sleepless nights staring out the window, chatting with AI about gravity, black holes, and theoretical physics. One night, I connected the dots with my major: How does gravity interact with mining?
I discovered gravimetry and was immediately fascinated by the idea of generating 3D models of underground mineral bodies without drilling a single hole. Upon researching, I found that the industrial software doing this costs tens of thousands of dollars. I decided to build my own. I originally thought this would take me 3 to 5 years and become my senior thesis, but AI changed the timeline entirely.
How I Built It
Since I am not a traditional software engineer, I had to invent a "Multi-Agent Orchestration" workflow using five different AIs to overcome my own technical limitations and token context windows. The engine operates on a strict loop: Prompt (\rightarrow) Validation (\rightarrow) Plan (\rightarrow) Execution (\rightarrow) Verification.
- Gemini acts as the Chief Architect for deep reasoning, context reading, and initial prompt generation.
- GPT/o3 takes that prompt for rigorous mathematical validation and system design.
- Claude Code reads the codebase, formulates a step-by-step execution plan, and implements the code.
- Codex/Antigravity environments handle the final physical and mathematical verification.
To maintain order, I created strict .md files that act as system guardrails for the AIs, and custom .nat files to isolate and trace bugs seamlessly.
What I Learned
I went from basic coding knowledge to understanding modern, cloud-native software architecture. I learned the critical importance of modularity—separating the mathematical Backend (Python/FastAPI/Polars) from the 3D Frontend (Next.js/Three.js) to avoid spaghetti code.
Scientifically, I learned how to translate theoretical physics into code. I dove deep into geophysical inversion mathematics, implementing concepts like Tikhonov regularization to solve the non-uniqueness of gravity: $$ \min ||W_d(Gm-d)||^2 + \lambda^2 ||W_m(m-m_{ref})||^2 $$
Challenges I Faced
- Token Limits: The biggest initial bottleneck was exhausting the context window of a single LLM. I solved this by engineering the 5-AI distributed workflow, allowing each model to handle a specific micro-task.
- Computational Scale: Transitioning from simple synthetic demos to an industrial-grade engine was brutal. Calculating gravity for large voxel grids caused memory overloads (OOM). I had to learn and implement HPC techniques, sparse matrices (CSR), and Apache Arrow (Zero-Copy transport) to render over 100,000 3D voxels in the browser without crashing.
- Mathematical Truth: Ensuring the engine didn't just generate "pretty visualizations" but actually respected geological constraints and physics was the hardest, yet most rewarding, challenge.
Log in or sign up for Devpost to join the conversation.