AI Enthusiast
From R Programming at Johns Hopkins to ML fraud detection at Amazon to bringing AI into automotive reconditioning at Carvana — pragmatic, measurable, and human-centered.
By the numbers
A snapshot of AI/ML work at production scale.
100k+ TPS, in production
Real ML work, at real scale, against real adversaries.
"Contributed to the design of a services-oriented architecture for the collection and identification of risky sellers operating at over 100k transactions per second. Product-managed a 14-FTE research and applied-science team through the creation, onboarding, and deployment of risk-detection signals stored in a graph database to detect and enforce multiple fraud rules and ML models."
Applied AI work
Where AI/ML met a real business problem.
AI in automotive reconditioning
Introduced ML models and AI tooling into an industry that had been running on '90s-era processes — applied to inspection, routing, and operations across a 42-site reconditioning network.
Worldwide fraud detection
Multiple fraud rules and ML models deployed across 12+ marketplaces, with risk signals stored in a graph database. Automated enforcement actions grew 18% YoY.
Visual seller-relationship app
Worldwide app to visually identify seller relationships among bad actors. Investigations per hour rose 87% — a force multiplier for the human analysts in the loop.
Operations & reporting BI
Department-level reporting consolidating dozens of metrics into a single visualization deck — task time reduced 35%.
HD Project Intelligence
Self-service BI cloud platform for industrial sectors — analytics, third-party tech acquisition, and a unique multi-segment business model.
Coding alongside AI, daily
Working hands-on with Claude Code, Roo, Cline, and ChatGPT to ship internal tooling, prototypes, and this very site — building intuition for where AI helps and where it doesn't.
AI / ML / Data stack
What's on my desk today.
AI capabilities — strength & depth
Self-assessed across the day-to-day stack.
Philosophy
How I think about applying AI to real product problems.
Pragmatic over flashy
The right answer is often a rules engine plus a small model — not the largest model that fits the GPU budget. Solve the business problem first.
Measurable outcomes
Every AI feature should ship with a metric attached: enforcement actions per hour, investigations per analyst, NPS, time-to-decision. If you can't measure it, you can't improve it.
Human in the loop
The 87% lift came from a tool that made human investigators faster — not one that replaced them. The best AI products augment expert judgement.