Projects
Research code, engineering tools, and a few things made for fun. Some of it is private to a lab or an employer; those entries are described rather than linked.
Efficient retrieval for planet-scale image geolocalization. Instead of encoding a query against every tile, the gallery is organized hierarchically and a learned model decides which regions to refine, so recall matches exhaustive search at a fraction of the cost. Name and details withheld while the paper is under review.
Location encoder built from learnable spherical Voronoi partitions, with shared global semantic tokens distilled from satellite imagery. Released with code and pretrained weights.
Scan configuration and link-budget tool for digital-code-modulation radar. Models theoretical SNR across range, Doppler, azimuth and elevation from simulated or measured antenna patterns, with a Dash GUI. Its curves became the basis for production test limits and customer coverage maps.
Data platform behind university research-strategy diagnostics: collectors for NSF HERD, OpenAlex, USPTO and NIH RePORTER, peer benchmarking, topic clustering of a campus’s output with SPECTER2 embeddings, BERTopic and Leiden communities, and report builders that emit LaTeX and PowerPoint.
Converted HL7v2.x lab results to FHIR resources and joined them with PracticeFusion EMR records into PostgreSQL and MongoDB; Spark ETL consolidating labs, billing and patient records into a queryable data lake with Apache Superset on top.
Arduino-based control and timing system for training animals on behavioral tasks, with a custom PCB and Python/C firmware. 85% cheaper and 87.5% lower latency than the commercial system it replaced. Presented at SfN and ABRCMS 2019.
Real-time dashboard for patient locations and status inside a hospital, with a 3D ward view in Three.js and a Node.js backend serving patient data over REST.
Reinforcement learning for Rocket League: PPO agents bootstrapped from human replay data, built for the deep RL course with SeongJun Kwon and Edwardsen James.
Interpretability experiments probing how language models represent arithmetic and mathematical reasoning internally, with evaluation harnesses for validating model behavior.
A local GPT-2 training run on FineWeb: data preparation, tokenization, training loop and evaluation, end to end, to understand what the tutorials leave out.
A chatroom and mock forum deployed on AWS (EC2 for the app, S3 for static assets) and again on GCP, mostly as an excuse to learn both.