I'm a full-stack machine learning engineer that likes solving real-world problems. Currently, I'm at Modal, a serverless compute platform free of config files. Previously, I've worked at Edlight building student grading solutions for teachers with fine-tuned models and Procurement Sciences building proposal writing tools for government contractors.
When I can, I build open-source projects and participate in competitions I think are cool. Some highlights:
- An attempt to clearly show the reach and limitations of studies when applied to lifestyle choices.
- An exploration of how Qwen2.5-0.5B multiplies using mechanistic interpretability techniques.
- Winner of the GPU Mode practice round.
- An ultrasound substructure localization system as the final project for SCU's ENGR 110 course and a POC of LLMs for pixel-level vision tasks.
- A minimal implementation of flash attention in CUDA alongside a visualization of its effect on input embeddings.
- formless, a hard handwriting image OCR system.
- admirer, a CLIP + GPT-2-based VLM built nearly two years before any major provider release.
- chexray, a chest x-ray diagnosis and report-writing system that beat contemporary SOTA models with a fraction of the data.
In my free time, I'm a climber. I've redpointed up to 8a+ (5.13c) on lead and 7C+ (V10) in bouldering. I've also onsighted up to 7c (5.12d) and 7B (V8). I also enjoy learning more about health, cooking/baking, and reading.