Services

Four kinds of work account for most of it, all four unremarkable. Their value is in the hours that stop being spent.

Connections between systems that run on spreadsheets and e-mail today

Two programs do not know about each other, so somebody retypes. I build the interface between them, including the part most people leave out: what happens when one side does not answer.

Reading and sorting documents

Delivery notes, incoming invoices, test records, orders that arrive as a mail attachment. A model reads them, the doubtful case goes to a person. That includes a set of evals measuring how often the model gets it wrong. Without that measurement the step cannot be signed off.

Workflows with an approval step

The queue in front of a signature is the longest leg in almost every survey. What can be shortened is not the signing itself but the path to it: the case arrives complete, with everything the decision needs.

Reporting nobody assembles by hand any more

Figures from the ERP, time recording and the warehouse at one cut-off, daily, without somebody spending a morning on it at the start of each month.

I work in what you already run: ERP or merchandise management, time recording, warehouse, file storage, mailbox, and whatever else accumulated over the years. Which product that is decides what the integration looks like and how long it takes, but not whether the rebuild is possible. The survey settles that before I name a price for the rebuild.

Where there is an in-house IT team, I work in their repository, through their review and their deployment. I take nothing away from anyone: your IT should carry the code on, not inherit it. Where there is none, I build so that an external provider can take over without asking me.

Where the data sits is your decision. If a model has to stay in the building, it stays in the building: that costs more compute and saves you explaining yourself to your own customers. If it does not have to, a bought-in model is usually cheaper and better.

“We will integrate your systems” is easy to say. What the sentence is worth depends on whether somebody has done it against a system they do not own and whose vendor does not call back. Four examples, all four out of production:

Four POS systems, one interface

In a pharmacy platform, Oracle SQL, IBM DB2 over ODBC, a vendor REST interface and a vendor web service all sit behind a single read-only model. The Oracle schema had to be partly reverse-engineered; the DB2 connection was debugged on site in the pharmacy.

A national toll infrastructure

An online shop binds every ticket sold to a licence plate through the national video-toll system — six toll operations against a formal three-party interface specification, across several versions.

Exchange, without opening Exchange

Delivering calendars out of a corporate Exchange without the web platform outside getting any access. For the second client the same connector ran against a different Exchange generation after about ten days.

A national system, reconciled

Importing courses from a national organisation’s central system, reconciling fully — insert, update, delete — and exporting registrations back on a schedule. In production for years.

In all four cases the effort sat not in the foreign system but in the failure case: what happens when the other side does not answer, answers twice, or changes its format in the middle of the night. Half the effort goes there.

The most valuable part of a rebuild like this is almost never the technology, but the knowledge that sits in your people’s heads and in your spreadsheets and nowhere else. The work consists of writing it down so that a machine executes it.

Carpentry rules as machine data

A configurator splits an overall dimension into carcasses that are actually producible, structurally and on the machine, and emits one machine project per carcass. Those rules previously lived in people’s heads and in price tables.

Funding rules as a rules engine

The calculation sheets of an education grant were translated one for one into a configurable rules engine: income types, thresholds, per-qualification calculations. Configurable, because the rates change every year.

Lab data as a public answer

A test lab’s raw measurements become publicly comparable results, resting on pre-aggregated tables so that a multi-year comparison stays fast.

Just as important, and harder to find on most providers’ sites:

I do not sell websites, online shops or SEO

An agency does that better and cheaper.

No licence reselling

I take commission from no vendor. Whatever has to be bought, you buy directly, on your terms and your contract. So I can also tell you when an off-the-shelf product is enough and a rebuild unnecessary.

No training as a separate product

Bringing your people onto what I built belongs to the handover and costs nothing extra. A general seminar about artificial intelligence you can get elsewhere.

No assessment without measurement

I do not offer a potential-assessment workshop where things get estimated. Put the same time into one day of walking the process and you end up with figures.

No work without access

I can only stand behind software I put into production myself, and that needs write access to the production systems. With an audit trail, with review, and with a way back for every change.