What I hold fixed
Remove before adding.
Most systems I see suffer from what was added, not from what is missing.
In practice: a diagnosis from me always includes what I would cut, and when a simpler model, a simpler pipeline or no model wins, that is what I recommend.
Fix the cause, not the symptom.
A patch on a surface fix becomes the next thing to debug.
In practice: when training and serving disagree, I trace the mismatch to the transformation that causes it, fix it there, and freeze it with a test so the same class of bug cannot come back quietly.
One thin path that runs, before deep machinery.
In practice: the first thing I build is the whole path — data in, decision out, a number someone can look at — with every stage naive. Then I improve the stage that limits the result.
Keep the uncertainty.
A single score hides how much you do not know. A draft that looks nearly done is not nearly certain to be finished.
In practice: I compare candidate models on expected business impact with its spread, not on a point estimate, and I promote a challenger through stages instead of in one switch.
Decide what "correct" means before building.
Objective first, then the invariants that protect it, then code standards, tests, performance and implementation. When two conflict, the higher one wins.
In practice: model output goes through a typed contract that is written before the code. The model may vary; the contract may not, and a broken contract fails the change instead of the demo.
A gate is only a gate once I have seen it fail.
A check that has never gone red is decoration.
In practice: when I add a quality gate to a pipeline, I plant a known defect and confirm the gate blocks it before I trust it with real changes.
Evidence before claims. Cite or omit.
A number typed into prose is a bug even when it is right.
In practice: I keep a register of every public claim about my own work, each one graded backed, qualify, contradicted or unsourced. This page uses only backed ones.
The reviewer is never the author.
Agent-written work is reviewed by a model from a different lab, because models from one family share blind spots.
In practice: a review comes back as failing cases — an input, the expected output, the actual output — not as adjectives.
Agents carry the volume. Judgment stays with a person.
Agents write code; I decide what counts as correct, and I read results, gates and risks rather than every line.
In practice: each change lands through an automated gate; only high-risk changes get a human read.
Your attention is the bottleneck, not my hours.
Option menus are a way of handing my work back to you.
In practice: I come with a decision and the argument for it, and I only ask the questions that only you can answer.
You own everything, from the first day.
Open source and local first where it is good enough; no vendor you cannot leave.
In practice: code, prompts, evaluations and gates live in your repository, so the speed stays when I leave.
Specialist in the method, generalist in the domains.
The method — measurement, evaluation design, calibration, causal reasoning, and the delivery harness around them — is what I refuse to be average at. Domains are where I have applied it: ad-tech ranking and user acquisition, fraud and risk, LLM evaluation. Stacks and model generations are tools; I adopt whatever works and defend none of them.
In practice: I will tell you which part of your problem I am strong at and which part I would hand to someone else.