About
Turning frontier AI capability into systems people can actually use.
I work at the boundary between model capability, product judgment, and deployment reality: the layer where AI systems have to fit real workflows, expose evidence, preserve human judgment, and survive contact with users.
The interesting work is no longer just proving that models are powerful. The hard part is building systems that create dependable value in the environments where people actually work.
Current thesis
Frontier models are making capability cheaper.
The bottleneck is shifting to workflow fit, trust, evaluation, and distribution.
The next valuable AI products will not just be better chat boxes. They will be systems that understand real workflows, preserve human judgment, expose evidence, and improve through tight feedback loops with users.
That is the area I am exploring.
Current experiments
My public projects are small tests of the same broader thesis: frontier AI becomes valuable when it fits real workflows.
ATO Copilot
Source-backed AI for high-trust review workflows. Tests how AI can help humans move through compliance and authorization work by preserving evidence, provenance, and accountability.
RFP Map
Public procurement data as a market-sensing interface. Tests how messy government opportunity data can become terrain to explore, cluster, and reason over.
Bare-Metal AI Lab
The infrastructure layer underneath AI products. Tests serving, latency, telemetry, quantization, service replacement, and recovery below the API layer.
The common thread is operational translation.
Models are powerful. The hard part is turning capability into systems people can actually use, evaluate, and trust.
Writing
Notes on AI systems, product judgment, and high-trust workflows.
Optimize Against the Real World
Shipping early, watching users, and tightening product feedback loops before optimizing the wrong thing.
A Model Is Not an Operational System
Why model capability is only one layer of useful AI products, and why reliability, evidence, interfaces, and evaluation matter.
ATO Copilot and the Compliance Gap in Agentic Software
Why high-trust workflows need source-backed systems that keep humans accountable.
What I believe
- Capability is becoming abundant. Workflow fit is still scarce.
- A model is not a product. A product is a workflow that changes what happens next.
- The best AI products will learn from real users faster than competitors.
- High-trust systems need evidence, evaluation, and human accountability.
- The fastest way to improve judgment is to ship, watch, and update.
Proof base
At Google Public Sector, I work on frontier AI systems for high-trust environments. Before Google, I built large-scale ML and cloud platforms at Capital One, including model-training infrastructure used by thousands of data scientists, machine learning engineers, and analysts.
My research background is in deep learning, computer vision, and medical image analysis. At Vanderbilt University, I developed new ML algorithms and published peer-reviewed research.
That background shapes my engineering taste: models matter, but deployment, observability, data quality, and human trust determine whether they create value.
Open conversations
I like talking to people building at the edge of:
- frontier AI and real workflows
- high-trust systems
- public-data interfaces
- model evaluation and operational reliability
- AI products where the hard part is workflow fit, not model capability
If that overlaps with what you are building, I am always interested in comparing notes.
Explore: Experiments · Writing · Labs · GitHub · LinkedIn