Independent AI engineering practice / Sonoma County
AI systems,
zero to one.
I take AI systems from first prototype to production you can measure. Get the data and schema right, evaluate what the system actually does, and feed those evaluations into every change.
FIG. 1 / THE PAA LOOPpaa.dev
01
Data layer & schema
trace { input, output, model, cost, grade }
every run captured, typed, queryable
every run captured, typed, queryable
02
Evaluation
03
Feed it back
evaluations inform the next change,
and whether more autonomy is warranted
and whether more autonomy is warranted
returns to 01 with a better schema
02 The work
All work and studies Measured, then shipped.
CONTENT CLASSIFICATION · RUBRIC REDESIGN Separated scoring from policy. Agreement up on a blind holdout. 64% → 77% keep/drop agreement CODING MODELS · PAIRED EXPERIMENTS Found what drives helper reuse in model-written code. Helper visibility > duplication AGENT PAYMENTS · X402 + ERC-8004 Paid API access for agents, priced by onchain reputation. Reputation unlocks discounted routes
03 The approach Data first.
Data first.
Then numbers.
Then progress.
Every engagement follows PAA, the 0 to 1 AI framework. You get a system you can measure, and a team that knows how to keep improving it after I leave.
Read the framework at paa.dev- 01 Structure the dataDefine the schema for what your system sees, decides, and produces. Capture every run so it can be queried later.
- 02 Measure what it doesBuild evaluations against your real criteria, checked against human judgment, so quality is a number and not a feeling.
- 03 Feed it backEvaluation results drive each change, and each change is measured. Autonomy grows only where the numbers support it.
PUBLISHED STUDIES Methods, data, and results in the open. Built on Assay, Scout, and JIG.