Spent the weekend building AI systems that could potentially save lives at AITX's DGX Spark Hackathon in Austin. While my brain is fried, it's in the best way possible.
Here’s what 48 hours of caffeine and GPU power produced:
- Hotspot Agent Watcher: Developed a model to predict where car crashes will happen in Austin before they occur. It utilised historical incidents, weather data, and even UT football schedules, revealing that game days lead to traffic chaos. Adding rain data significantly improved prediction accuracy, highlighting the importance of weather beyond just peak traffic times.
- AutoHDR: Created a physics-based image enhancement system that operates like a professional photo editor. It’s fast, smart, and effective.
- Port-to-Rail Forecaster: Predicted shipping congestion days in advance using graph neural networks across the North American rail network. This means if the LA port gets overwhelmed, we can anticipate the ripple effect reaching Chicago before it happens.
This experience was not just about making code run faster; it was about understanding why GPUs function the way they do. This knowledge became clear during debugging at 3 AM.
The real breakthrough came from running hundreds of experiments to enhance accuracy. Without GPU power, we would have settled for "good enough."
Huge thanks to my teammates for an incredible weekend: Ryan Blake Story for the strategic architecture and keeping us focused when things got chaotic, Poorav Desai, Saroja Vuluvabeeti, Afnan Ahmed Crystal, and Stuti Patel for pushing through the 2 AM walls and making this happen.
Built with PyTorch, XGBoost, RAPIDS, FastAPI, Next.js, and Mapbox.
A special shoutout to Jake O'Shea and Michael Daigler for outstanding event leadership. The seamless coordination, responsive support, and thoughtful organization made all the difference. You created an environment where teams could focus on building instead of logistics—that's rare and deeply appreciated.
Thanks to AITX Community and Antler for hosting, NVIDIA, ASUS and Arm for providing world-class hardware, and City of Austin for open data.
Sometimes, the best learning occurs when you’re too tired to overthink things.