I build the wiring between business systems, so information moves without being retyped — and can still be traced afterwards.
A practice management system, an accounting package, a document generator, a registry, a mail server. What is missing is the wiring between them — so information gets carried across by hand, and the carrying is where the hours and the mistakes go.
The work is not replacing any of it. It is connecting it.
Systems that were never designed to talk to each other: practice management, accounting, document generation, e-signing, public registries. One workflow, end to end, without re-entry.
Most of the effort goes into the edges: a lookup returns nothing, a document is rejected, a counterparty answers three weeks late.
Public and third-party APIs — mapping data, company registers, published web content — aggregated into structured output an application can depend on.
The first request is never the problem. Rate limits, schemas that change without notice and sources that go quiet all have to be absorbed before anything downstream notices.
Where data residency or audit obligations rule out handing records to a third party, I deploy on private servers inside the EU.
Including the language model, on that same infrastructure rather than a hosted API — and logs kept in a form that can still be produced years later.
Client intake built against Dutch Wwft obligations.
ProblemChecks run by hand across a register, a sanctions list and a risk matrix.
SolutionOne workflow verifies, screens for sanctions and PEP, and classifies country risk.
ProblemThe call still needs human judgement.
SolutionIt drafts the assessment and the letter. It proposes; it does not decide.
ProblemThe reasoning has to be reconstructable years later.
SolutionEvery lookup and decision is logged. Self-hosted, EU model, records stay put.
One physical place, resolved across mapping, geographic and encyclopedic sources.
ProblemOne building, three sources, three names, pins 40 to 100 metres apart.
SolutionFour merge passes. All lean the same way: not convinced means keep them separate.
ProblemFuzzy matching quietly collapses places that are genuinely different.
SolutionNames merge only on exact normalised equality. Two libraries stay two libraries.
ProblemThe source data looked rich and wasn’t.
SolutionAn audit first: 2,059 buildings, 7 names, no heritage tags. Coverage was a bulk import.
A review queue one person operates, where some decisions carry legal weight.
ProblemToo large to review by hand. Too consequential to leave to a model.
SolutionMachine proposes, human confirms, machine applies. Proposals never touch what is live.
Problem“Approve” is ambiguous. Agreement, or a decision of your own?
SolutionEach verdict asserts one fact. Model verdict and human decision are stored apart.
ProblemDeclining a contribution owes the person a reason.
SolutionThe API refuses to record a rejection without a written statement of reasons.
You know the architecture and what it costs to build and to run, before any code exists.
We walk the process as it runs today and separate what is mechanical from what needs a person. Not everything should be automated. The scope is whatever survives that conversation.
The proposed system, the components it touches, and what it costs to build and to run each month — written down before any code exists. A reference point, not an opening position.
Built against your real data and your real exceptions, not a clean sample. Deployed to your own instance, with the credentials issued to you rather than held on your behalf.
Documentation covering how the system works, how to change it, and how to run it without me. Support continues afterwards if you want it.
Self-hosted where the requirement calls for it, EU infrastructure otherwise. Records are not routed through services the task does not require.
In the moderation lanes the boundary is enforced by removing the destination, not filtering the payload. There is no second provider to cross to, and a test asserts it.
Open, portable tooling. If the arrangement ends you keep the system, the credentials and the documentation — and it carries on running.
Model size, hosting and storage chosen against what the work demands. Over-provisioning is not a one-off decision — it is a bill that arrives every month.
Usually the same decision twice. Moving record generation behind the human confirmation cut cost and cut data collected, in one change.
A short description of it beats a general enquiry. Email is the reliable route.