I build intelligence that has to be right.
I run product, engineering, and data at iQpay, where machine decisions move real money at national scale. When one of those decisions is wrong, a family gets declined at the register. I build the systems that make those decisions, and the evaluation and judgment that make them trustworthy enough to run on their own.
I came to machine judgment through the study of human judgment.
I read Psychology and History at Harvard: how people reason and how they fail to, how beliefs take hold, how whole societies decide what is true and what is allowed. Building systems that decide turned out to be the same question from the other side. What does it take to trust a judgment, made by a person or by a machine? Evaluation, it turns out, is epistemology with a deadline. The instinct I rely on most in production, distrust a confident answer until it has earned belief, I learned from studying people long before I applied it to models.
I run product, engineering, and data. The whole technical organization, and the intelligence at the center of it.
iQpay is a payments platform serving health and social programs across thirty-one states, deciding in real time and product by product what each program covers at the register. I lead a distributed team of four engineers, set the technical strategy, and make the decisions that carry risk: what we build, what we automate, and what a system has to prove before it is allowed to touch a live payment network. And I build the core myself: the classification pipelines behind those decisions, the evaluation systems that keep them honest, and the controls that let them run unsupervised, at a standard of correctness that is measured, not assumed.
A payment network built only to refuse products can, in my systems, be trusted to allow them. I built the precision that unlocked whole categories it had never been able to offer, and a fresh-produce standard adopted by a state health authority, now live in production.
Decisions that move money cannot be watched by hand at scale. I build multi-model juries, calibrated evaluations, and release gates that make automated judgment reliable enough to run unattended, and honest enough to raise its hand the moment it should not. Read the note on trusting machine judgment →
Seven million products, unified from national retail feeds, live sources, and field data I gathered in person, into the single platform every eligibility decision and every invoice draws on.
0.88 to 0.94, and the model never changed.
Our eligibility classifier had stalled at 0.88 F1, and every architecture we tried landed in the same place. Before spending more on models, I audited the ground truth they were trained and judged against, and found that roughly a third of the labels contradicted one another. So I made the unglamorous call: freeze the modeling, rebuild the labels, and re-measure everything against the repaired truth. The same model came back at 0.94 F1, and the pipeline has run unattended in production since, with model disagreements escalated to a person instead of settled by chance.
The technology lead left. The pipeline never noticed.
When our technology lead departed, product, engineering, and data became mine at once, reporting to the CEO. The visible job was continuity: three hundred programs, a live payment network, and the four engineers who kept every one of them running while the org chart changed underneath them. That credit is theirs. The quieter job was the one nobody was asking about. I brought in outside attackers for the company's first external penetration test, fixed what they found, and put HIPAA gating in place. A platform that moves health-program money carries two kinds of risk: the failure burning today, and the one arriving in a year or five. The first announces itself. Someone has to go looking for the second.
Seventeen million subscribers, eight cities.
Before iQpay I built product at Turkcell, Turkey's largest telecommunications company: its streaming platforms, and, alongside three data scientists, a location-analytics model spanning 17.3 million subscribers in eight metropolitan areas that the company adopted for tiered pricing. I have never stopped reading in the seams between fields, behavioral science, philosophy, the history of ideas, where the questions worth asking tend to live.