Experience

I build and operate the retrieval and agent infrastructure behind customer-facing AI features. That covers RAG services in Python and FastAPI, multi-tenant agentic workflows exposed to internal platforms over Model Context Protocol servers, and the CI/CD that gets containerised services onto Kubernetes.

In practice it means owning things end to end: architecture, database performance, authentication and tenancy boundaries, release automation, and the observability needed to know whether any of it is actually working.

A ten-week Digital Transformation Summer Corps project. My team built a multimodal retrieval system that let public health and environment researchers ask questions about neighbourhood conditions (walkability, greenery, safety) directly from street-view imagery, rather than surveying areas in person.

Most of my time went to two things: working with non-technical faculty stakeholders to turn open research questions into evaluation pipelines we could actually measure against, and inference optimisation, without which the system was too slow to be usable. The work was covered by WashU’s The Source.

Infrastructure work underneath the lab’s research: distributed training pipelines in Python multi-terabyte image datasets on GPU clusters, plus the build systems and asynchronous data-ingest tooling around them. The recurring problem was that research code assumes one machine and one dataset, and neither assumption survives contact with a cluster.

I also collaborated on model work, including a depth-conditioned transformer submitted to WACV 2026.

Detection and segmentation of orthopedic hardware in medical imagery, using a Mask R-CNN with a Swin Transformer backbone. I worked on the architecture and on the regularisation needed to make it generalise beyond the training distribution, which is the part that decides whether a medical imaging model is useful or merely impressive on a held-out split.

I took it through to an interactive Gradio application on Hugging Face Spaces, trained on cloud GPUs, so people without a Python environment could try it.

Led computer vision for WashU’s robotics team: real-time object detection on embedded NVIDIA hardware, model export and quantisation for edge inference, and a CUDA-accelerated pipeline running from camera capture through to display. I also wrote the trajectory logic that used sensor feedback to aim an autonomous launcher.

This is where I learned what a latency budget feels like when it is set by physics rather than by a service-level objective. A frame that arrives late is not slow, it is wrong.

Hundreds of hands-on classes in robotics and programming for K-8 students, using LEGO EV3, Scratch and Arduino, plus community STEM showcases. In the later years I led a small team of instructors and helped scale the curriculum.

Explaining a for-loop to an eight-year-old is still the most useful communication training I have had.