I'm a high school student in Austin, Texas. I build machine learning systems for cars and robotics — and then I try hard to break them.

Most of what I care about lives in the gap between "this works in a notebook" and "this works on an actual road." A lot of ML looks great on a benchmark and falls apart the moment it meets the real world. The interesting part is finding out which is which — then publishing the answer, including when the answer is that my own model didn't work.

A research platform that detects potholes three ways — with a camera, with a phone's motion sensor, and with both fused together — plus an honest test of whether any of it survives contact with a real car.

• 94 benchmark experiments with frozen train/validation/test splits
• Recorded my own drives, then hand-labeled 266 detection clips blind — no idea which model flagged which clip — and did the entire labeling a second time to measure my own consistency (97% agreement with myself)
• My best model failed in the real world. In the lab, fusing camera + sensor won. On real roads, precision fell to ~0.10 and the fusion advantage disappeared — the camera half went blind on dashcam video.

I published the negative result anyway, with the code, the pre-trained models, and a privacy-scrubbed dataset of my own drives. The takeaway: combining two sensors isn't a free win — a system is only as strong as its weakest sensor in a new environment.

I released the dataset because good data is usually the real bottleneck in these problems, not clever models. Anyone can clone the repo, record a drive with their phone, and re-run my experiments on their own car.

Accepted to GARSEF (Austin regional science fair), 2026. Next: I'm drafting an open letter to Austin, Phoenix, and San Francisco arguing that cities should pool road-condition data from every self-driving fleet — my own results are the argument, since a model trained on one vehicle doesn't transfer to another.

Curieux Academic Journal, December 2025, Part 2, Issue 2, p. 589 · read the PDF

Can one AI teach another? I built a two-agent feedback loop inspired by backpropagation: a "coding" AI drove a virtual racecar in the F1Tenth autonomous-racing simulator and logged its crashes, while a second AI analyzed those crashes and rewrote the driving code — with no human in the loop. The dual-AI system substantially outperformed both a human-coded baseline and a single AI with no feedback.

Summer 2026 internship with Prof. Kishore Kuchibhotla's lab at Johns Hopkins, applying my AI/ML experience to their research on lucidity in dementia — the episodes where patients with advanced dementia unexpectedly regain clarity and communication.

I built two tools for the project:

• an AI speech synthesizer for modeling lucid events, and
• an audio analyzer that helps detect lucid events in patient recordings.

I'd like to open-source both once the lab's research is further along. I got into this through the computational side, but it's personal too — dementia has affected my extended family, and pointing these tools at that problem is the most useful thing I can imagine doing with them.

I build the real version of a thing, then try hard to break it, then give it away. The building is the fun part. The breaking is where you find out whether it was ever true — and it's the part most projects skip. Open-sourcing it is how the field actually moves: someone else gets to start from my failure instead of repeating it.

Python · PyTorch · scikit-learn · YOLO / computer vision · signal processing · FastAPI

📫 github.com/warstad11/pothole — code, data, and full results