One loop, every project
I use the same six steps for each project. Many AI agents do the build. Gates check the work.
Six steps
Select a lens. Then run one task through the loop.
Research
Read the field. Map competitors. Write a short research note before code.
Four ideas I build on
Each one came out of a real project, usually after something broke.
-
Data quality beats a bigger model.
My detector stopped improving when the labels did. I changed the product so that every estimator correction becomes clean training data.
See it in Blueprint Vision -
A wrong answer is worse than no answer.
When the system is not sure, it says so and asks a person. The dealer sees the confidence, not just a pick.
See it in Collectible ID -
Give agents small, revocable authority.
Spend limits, allow-lists and session keys that expire. Code checks the result, not another model's opinion.
See it in Agent Systems -
A check that cannot run must say so.
Every gate returns pass, fail or error. Error stops the line and alerts a person. It never counts as a pass.
See it in Harness Engineering
Where I started
I did not start from zero. Cognitive science gave me the questions; computer science and a bootcamp gave me the tools; business work taught me to scope and deliver.
Rice University, B.A. Cognitive Science, 2017.
Cognitive science asks how a mind represents what it knows. Early AI grew from the same question: symbolic reasoning, the first neural networks, and learning from examples. I studied those early representations and the trade-offs between them.
Computer science coursework at Rice.
I took computer science classes alongside my major. They are the foundation for the engineering I do now.
A nine-month AI and ML bootcamp, 2019-2020.
How models train, how loss and evaluation work, and why models overfit.
About five years in business.
Solutions engineer, account manager, and owner of my own business. I learned to scope work with clients and deliver it.
