MS Candidate, Health Care Informatics
University of San Diego · Expected January 2027
I build and test AI tools, mostly on open-weight models running on my own hardware. Seven years as an emergency department technician and three as a QA analyst at Riot Games taught me to care about two things in any AI system: what it actually does, and how we know.
Machine learning in healthcare, database management, healthcare and population health analytics, biostatistics, systems analysis and design, electronic health records. Capstone: Opt-In-Time. EMT certification, 2019.
Chat and agent frontend for DeepSeek V4 Flash running locally: reasoning traces, live GPU telemetry, context-window management, a sandboxed agent mode with 14 tools, tests and a reasoning-mode benchmark. Python, JavaScript, Bash.
A 1-billion-parameter model on llama.cpp routes tasks to specialist prompt bundles. On a frozen 24-task suite, it beat the same model alone on GPU (71% vs 67%, 4.2× faster, 5.6× fewer tokens) and lost on CPU-only hardware (58% vs 71%); the write-up reports both.
Fine-tuned text model rating games on 27 emotional ingredients. 5-fold cross-validation showed it barely beat predicting the average and that hand labels didn’t predict reviews. The ingredients now underpin StoryGen and Lifestone’s AI-guided design.
Home check-in system for discovering health trends, in my family first, then for anybody, anywhere: Kotlin kiosk app, FastAPI and SQLite server, React dashboard. Records missed check-ins on purpose, because bad days go unlogged.
Classified CKD stages from routine lab values in scikit-learn. Traced a perfect 100% score to target leakage, removed the leaking variable, and reported a realistic 86.8%.
EHR alert with three if/then rules for children who may qualify for the newly approved OTOF gene therapy before a cochlear implant closes that option for the ear. Review of 10 sources, requirement tests, evaluation plan and FMEA.
Cron-scheduled agents on local models: over 300 commits in ten days, and a fine-tuning dataset built from their failures.
Three.js engine and SpacetimeDB server built by parallel AI agents; each new check must fail once before it’s trusted.
Also: UIXtend (released screen magnifier, C#), AIWar (Rust server, ML training gym), OmniQuest (hand-written Godot roguelike).
Earlier: Windows Server and Linux administration, 2012–2013. Marketing consulting with Power BI reporting and SEO.