Jordan Beary
I am a data scientist and analytics manager. For the past eight years I have moved a growth-stage company from guessing to measuring: I founded and lead its analytics and data science department, and my daily work is translating business problems into data science, and data science into decisions that hold up in front of both executives and engineers. I have taken new business verticals through the full data maturity model — collection, reporting, prediction, automated optimization — and my domain is marketplace and auction mechanics in consumer lending: real-time bidding, lead routing, fraud, experimentation. I have spent ten years building the analytics infrastructure that work runs on.
I hold an M.S. in Systems Science, where I studied applied and theoretical modeling of systems. This site presents one project in depth: Tributary, an end-to-end case study that takes a simulated lead marketplace’s business questions and turns them into decisions backed by identity modeling, calibrated models, and an honestly read experiment — directed by me, built with AI agents, with the direction itself kept as an auditable record.
Three proof points
Decisions from data: a reserve policy worth about $4.2M a year. On a simulated lead marketplace, replaying 2.4M logged auctions under candidate reserve schedules showed the floors were set at the right level but the wrong shape: a stiffer tier-1 reserve with much cheaper deep tiers earns +2.02% revenue per lead (95% CI +1.96% to +2.08%), about $4.2M a year at the logged volume. Re-running the simulation engine across five fresh worlds realized +2.98% (range +2.88% to +3.15%), and a stress test put the break-even where buyers give back about 19% of the tier-1 increase — so the recommendation ships staged and randomized, to measure that response before the full schedule goes live. The same unified view sized a separate open question: $12.0M, 5.7% of revenue, that buyers paid for consumers they had already bought within the prior 30 days. The strategy memo → · The findings →
Identity as a measured model. Three systems with no shared key were resolved to 635,579 consumers by probabilistic linkage where identity is fuzzy and deterministic linkage where an exact key exists, clustered at the high-precision operating point because transitive closure amplifies every false edge, and scored against a ground-truth crosswalk the pipeline never sees: link F1 0.879, dedupe F1 0.873, cluster purity 0.9998, and 95.3% of auction events joinable, from 0%. Identity modeling →
A method you can audit. Every phase shipped with decision records, an intervention log of agent errors and the corrections that fixed them, and per-commit provenance separating human direction from agent execution — a harness for directing AI agents, kept so an evaluator can decompose who contributed what. How this was built →