Method
Diagnose, fix, prove, hold.
The same four stages on every engagement. Each one ends with something you own, and you can stop after any of them without owing the next.
01
Diagnose
5 days
Findings in writing, ranked by severity, each with the cost of leaving it alone. Fixed price. You own the document whether or not we continue.
02
Fix
2–6 weeks
Agreed scope, fixed price, explicit definition of done. No hourly billing, no scope drift, no surprise invoice.
03
Prove
2–4 weeks
Before and after, measured. If the numbers did not move, that gets said out loud rather than buried in a status deck.
04
Hold
monthly
Monitoring, alerting and a monthly review, so the thing that was fixed stays fixed. Optional, cancel any month.
Four things you can hold me to.
Not values on a wall. Each one is something you can point at mid-engagement and say I did not do.

Evidence before opinion
I read the code, run the command and check the output before telling you what is wrong. A diagnosis from a conversation alone is a guess in a suit.
Fixed price, fixed scope
Hourly billing rewards the wrong thing. You get a number and a definition of done before work starts, and if scope changes you get a new number rather than a larger invoice.
Bad news travels fastest
If something is going to slip, or a measurement came back worse than hoped, you hear it that day. Status reports that only contain good news are how projects arrive at their deadline broken.
You keep everything
Source, documentation, infrastructure, reasoning. No dependency on me by design. If the engagement ends your team can carry it forward without a call.
How the work gets done
The draft got cheap. Being right did not.
A first version of almost anything is now cheap to produce. That is real, and pretending otherwise would be dishonest. What did not change is the cost of knowing whether that first version is correct — and that cost went up, because there is far more plausible-looking output to be wrong about than there used to be.
I have taken three systems into production alone in the last eighteen months. In each one the distance between the demo that worked and the version anyone could depend on was two to four months. The demo got cheaper. That distance did not move at all.
That gap is the subject of a longer note, written from those three.
Measured, not assumed
Where software built this way actually fails.
Not risks in the abstract. These are failure modes I have hit in my own systems and measured, which is why each one has a number against it and a control behind it.
zero callers
Complete, correct, and never called
Scoring code sat in one of my own serving paths, fully written and compiling, with nothing invoking it. Compiling is not integrated, and integrated is not correct. It is now the first thing every review I run looks for, because the version of this that reaches production is a safety check nobody knows is switched off.
4 phases “done”
A green tick over something that never ran
An agent reported four phases complete. The interface it had built had never once rendered. Self-reported progress is the least reliable signal available, and it is the easiest one to build a schedule on.
Six defects behind a green tick1 in 2 → 2 in 3
The shape fits, the meaning does not
Constraining generation to a schema removes malformed output entirely. It does nothing about output that is well formed and wrong. On a golden set, parsing succeeded every time once the schema was enforced, while first-pass semantic validity was half, rising to two thirds after tuning.
1 in 6
Invalid before anything catches it
Where a model produces something another system consumes, roughly one draft in six came back structurally invalid before a repair step existed. A generation step that is quietly wrong one time in six is worse than not having one at all, unless something stands behind it.
2 of 2
A number that was never true
In two prose generations out of two, a model invented a statistic — stated in the same tone as the true sentences around it, once with a plausible source attached. This is the most expensive failure of the set, because it is invisible at review unless the reviewer already knows the answer.
When the model is right and the number is wrong
The controls
What stands between a draft and your repository.
None of this is a product you can sign up for. It is the machinery that makes a five-day diagnosis possible and keeps a fast first version from reaching you as a finished one.
The procedure, not the recollection
Around thirty review procedures — access control, data flow, business logic, database behaviour, performance, security — written down as executable checklists rather than held in my head. A diagnosis runs the same way every time, and the findings document is the output of the procedure rather than a recollection of one. An area nobody examined appears as examined by nobody, instead of as silence.
Evidence gathered by machine, judged by a person
Internal services gather what would otherwise take a week: codebase surveys, dependency and version drift, support triage. They produce evidence. They do not produce conclusions, they decide nothing, and none of what they return reaches you without passing through someone who can be argued with.
Nothing generated reaches a system unchecked
Where a model produces something a system consumes, generation is constrained to a schema, the result is validated for meaning, and a repair step runs before anything downstream sees it. The schema catches shape, the validator catches meaning, and both are needed — because the first one passing is exactly what makes the second one easy to skip.
A fact is written by someone who could have written it anyway
Structured output is safe to automate. Prose asserting a fact is not. Anything claiming something true about your system, your numbers or your users is written or verified by a person who could have produced that claim unaided. It is the only check that survives a model being confidently wrong.
Your code does not have to leave my machines
I run open-weight models on hardware I own, including a thirty-billion-parameter model for text. Work that should not leave your infrastructure does not have to sit in someone else's either. Where a hosted model is genuinely the right tool I will say so, and you decide.
Separately from any of this, building AI systems for clients — constrained generation, routing, evaluations in their own products — is a capability in its own right, written up under AI-native development. The agent side is being built out as a product under the Layap name — Layap Agent Platform. It is early and currently a waitlist rather than something you can log into, so treat it as a look at the direction, not as one of the three systems above.
Answerable
Who carries it when it is wrong.
The number is agreed before the work starts
Fixed price against fixed scope. If I have misjudged the effort, that is mine to absorb rather than yours to discover on an invoice. Scope that changes gets a new number and a new date, in writing, before anything is built against it.
Defects are my cost, not your next project
Work that does not meet the agreed definition of done is re-performed at my expense. A warranty is cheap to offer and expensive to honour, which is precisely what makes it worth reading as a signal.
Someone else drives at the handover
The handover is finished when an engineer you nominate deploys the system and restores it from backup using only the written material, while I watch and do not touch the keyboard. If they cannot, the material is wrong, and fixing it is my job rather than their training problem.
Fifteen years underneath it
Three production exchanges designed, built and deployed alone. Nine products taken from idea to launch. The judgement being applied to your system was not acquired in the last two years, and it is the part of this that no tool supplies.
If I billed by the hour, a faster draft would be a discount you were entitled to ask for. I do not bill by the hour. You are paying for the outcome, and for someone carrying the risk of being wrong about it — and that risk did not get cheaper.
Start at stage one.
A teardown is five days and a fixed fee. You get the document either way, and no obligation to continue.
Get in touch