Scaling AI is not a model problem. It is an engineering, governance, and trust problem. I worked across banking, insurance, and telecom — industries where the context changes, but the failure modes often rhyme.
Research tells me what to design for.
Engineering makes it scalable.
Governance decides what survives production.
Nineteen years putting technology into places where it is not allowed to fail — banks, insurers, telcos. There, the interesting question is rarely can this work. It is will anyone let it run.
A model that performs well in evaluation and a model that survives a risk committee are different artifacts. I work across the distance between them.
I build. Data engineering: pipelines, models, and lineage that make a platform trustworthy where the data enters, not where it is presented. AI engineering: retrieval, orchestration, evaluation harnesses, and the instrumentation that shows what a GenAI system is doing at scale — not what it did in a demo. Governance enforced in the system, not described in a document.
I research. I hold a PhD in Artificial Intelligence and continue to publish on responsible AI, machine learning, and model behaviour under sustained use. My recent work examines how sycophancy compounds across long conversations — where agreement at turn three quietly conditions agreement at turn thirty. These failures are invisible in a demo. They surface only once a system has been live long enough to matter.
Production tells me what to study. Research changes what I build next. Most people run one half of that loop.
What breaks between a working demo and 100,000 users.
Frameworks that get models past risk committees and into production in regulated environments.
Analytics that change decisions, not dashboards that describe them.