Data Scientist, Uber · Doctoral Praxis Advisor, George Washington University · Doctor of Engineering · U.S. Air Force veteran
I build the analytics products and AI workflows that leaders make decisions on. I test AI systems on real work and publish what held up, what failed, and the evidence behind it.
Background
I’ve spent my career making data safe to act on. In U.S. Air Force accounting and operations, I made it a rule to reconcile before I report. As a Business Intelligence Engineer at Amazon Web Services, I owned the reporting behind a managed-services business’s weekly operating review and its 24 customer-delivery managers: eight dashboards that cut ad-hoc reporting requests ~80%, a monthly finance reconciliation taken from about three days to three hours, and an 8-hour cleanup reduced to about 20 minutes.
At Uber, I own the executive reporting portfolio for spend pacing and performance across marketing channels and multi-billion-dollar verticals; I’ve cut reporting time ~40% and hold my reporting products at 99.9% uptime. I led the consolidation of fragmented reporting into one governed analytics product that catches definition drift and deployment failures before users hit them: metric-level comparison before cutover, a rollback path, and repeated data-quality failures turned into automated gates. I’ve also built the usage measurement that turns engagement into keep, improve, consolidate or retire recommendations, and a brand scorecard that saves one team about 100 hours a year.
On the AI side, I reduced monthly and quarterly business-review deck drafting to one click with an AI automation that refreshes the slides and writes data-grounded performance narratives through layered queries and review. I’ve built the monitored, failure-aware integration and review workflow that brought an AI service into a live product, so failures stay visible and a person still makes the call. Before trusting a lower-cost local model, I designed a replayable evaluation, with fabrication and abstention tests, that showed where it was usable and where its reasoning fell short.
At George Washington University, where I advise three doctoral candidates, the question I press hardest is whether the evidence supports the claim. AI deserves the same question: a fluent answer can be wrong, stale or impossible to reproduce and still sound certain.
Selected work
Open source · Python
A job that goes quiet raises no exception, so error monitoring never fires. My detector watches the gap since the last output instead. On a 105-case synthetic benchmark it scores ~84% accuracy (~82% recall, ~86% specificity), with every error in the borderline cases I built to be hard.
Live product · synthetic demo
I defined the product and, with AI coding agents, built it: resident, manager and owner views, plus a local-model assistant that receives only the records the signed-in user may see and cannot change them. During hardening I caught signed-in-only contact data shipping in public JavaScript, moved it behind an authenticated server route, and extended the privacy scan to the compiled files.
Evaluation design
The AI manager designed, ran, graded and won the first run of my task-difficulty benchmark, so I discarded it. I rebuilt it sealed: I hashed the manager’s difficulty estimates before any worker ran, froze the answer keys and graded by script. Nine of the 11 gradeable cases came back harder than predicted.
Home lab
My home-lab benchmark first rated the local model worse than a paid one. I traced the fault to my own pipeline: it fed born-digital documents to every model as images. Once I switched the inputs to text, all four models, two local and two paid, scored between 90% and 94% in small runs.
Credentials
For my praxis, Optimizing Robotaxi Deployment Through Behavior-Based Demand Clustering: A Case Study in Urban Mobility (published via ProQuest, 2025), I clustered a ~70K-record dataset derived from public Boston ride-hailing data by rider behavior, validated the clusters with Kruskal–Wallis tests, benchmarked them against geographic segmentation, and turned them into hub, fleet-mix, staging, scheduling and pricing recommendations.
How I work
01
I put AI to work on property operations, civic reporting and financial records, where a missing or stale record costs someone.
02
I freeze answer keys before a model runs, and the model that produced an output never scores it.
I’d like to hear from people building, evaluating or leading applied-AI work, and from anyone who thinks I got a result wrong.