The expertise is already on staff.
The engineering isn't.

We build the system that gets what your specialists know out of their heads and into software you can question.

For research institutes and expert organisations: a dozen people who know the field, no engineering function, and one person everybody quietly routes the hard questions to. Everybody also knows roughly what year that person retires.

Arrange a 20-minute call

Five figures a month, bought a week at a time. You can stop after any of them, and one client bought forty weeks that way.

Ush, a UK executive-scheduling company. Different sector, different build, not a research engagement — we won't dress it up as one.

Forty weeks bought one at a time is thirty-nine chances to leave that nobody took. Better than a testimonial.

An agency asks for a specification

Working software in week one. The specification comes out of it.

Hiring takes months

Work starts next week. Nobody for you to interview.

Both hand it over at the end

Repository and cloud accounts in your name from day one.

Evidence

Their expertise was the requirement, and nobody had ever written it down.

A research institute had built out its domain knowledge thoroughly and had no route to hiring engineering. We became their engineering function for a team of fifteen doctoral scientists. Papers arrive as PDFs and their reference lists; a model API is called stage by stage to extract who claimed what, on which page. Ask the graph a question in plain language and the answer is the quoted passage and a citation to the page it came from.

Extraction a reader can check, rather than a paraphrase they have to trust. Within days of the pipeline existing, the scientists were producing knowledge graphs themselves. Delivery was accepted by their own domain expert, and the engagement is still running.

The institute is in Israel. Naming it needs their written permission and we haven't asked. You'll be introduced to them, by name, before you commit to anything.

Schematic — shape only

The line reads left to right: papers arrive as PDFs and their reference lists; extraction calls a model API stage by stage and writes triples that keep their page references; the graph holds who claimed what and where; a query API takes plain language and returns the quoted passage with a citation. Nothing untraceable ships, so a wrong extraction stays findable by construction rather than by spot-checking. The dashed box below is the platform that hosts a graph of that shape for any corpus. It exists, and it is not what answers questions today.

For the engineer this gets forwarded to

RDF triples in Apache Jena Fuseki alongside vector embeddings in Milvus, behind a SPARQL API and a hybrid semantic-search endpoint. Kubernetes-native, cloud footprint in Terraform, blue/green publish with rollback. Automated testing and containerised deployment went into the research codebase alongside the pipeline.

Method

Nobody has ever answered "what do you want" with a specification.

That isn't a failing on your side. It's what tacit knowledge is. So we don't ask your specialists how the software should work — that overconstrains the problem, and they don't know. We ask what makes the result good, then find an architecture that delivers it.

Week one is a running system with a first slice of your material in it, usually wrong enough to be a little embarrassing. That's the mechanism rather than a failure of it. Their corrections are the requirements process, and the specification accumulates instead of being guessed at the start.

Eight weeks of one engagement — shape only

What is running

The specification, so far

Every line on the right began as somebody saying "no, that's wrong, and here's why". None of it was written down before week one, because none of it could have been.

The work

How a week is possible at this price

The fleet builds and an architect decides. Most of the volume is done by language models under a harness we built and supervise; none of the judgement is. Nothing reaches your repository unsteered.

You'd have worked this out from the price eventually. We'd rather you heard it here. Before work starts we write down what the fleet may read, what it may write to, and what never leaves your accounts.

Schematic — shape only

The fleet hands work to an architect, who sends most of it back and passes the rest through. A delivery manager sets when the week lands, never what is in it. This is the difference between an engineering company that is agent-driven and a prompting tool: architecture, review, and somebody who answers on a fixed day.

Terms

Eight lines, and none of them need a lawyer to read.

Cadence: Working software weekly, with a recorded walkthrough the day before each call.
Acceptance: A written definition of done for each week, agreed before that week starts.
Continuity: The architect who scopes the build on the first call is on your weekly call and answers for what ships.
Cancellation: After any week. No notice period, no setup fee, no minimum term, and no exit conversation with somebody whose job is to keep you.
Custody: Repository and cloud accounts in your organisation's name from day one, not transferred at the end.
Your time: About an hour a week from each specialist, spent with the people building it rather than an account layer.
An estimate from earlier engagements, not a measured average. If your specialists can't find that hour the method doesn't work, and it's better to know now than in month two.
Included: Hosting, deployment, pipeline, automated checks, security testing, compliance groundwork. No infrastructure line item.
Hours: 08:00–17:00 CET, remote.

Fit

Read both lists before you book.

This is for you

Your specialists know the field deeply, and none of it exists in a form software can use.
Nobody on staff could specify, hire or assess an engineering function.
The budget is approved or the funding has landed.

This is not

You want a quote against a frozen scope. The method scopes by building.
You already have technical leadership. We're an engineering function, not an addition to one.
Your specialists can't spare an hour a week.
Nobody on your side knows the domain. We can build it; we can't know what it should do.
You want a tool your own team operates. Different purchase, wrong vendor.

Two limits aren't negotiable: delivery is remote with no on-site placement, and your team has to sit within two hours of CET, because the weekly call is the mechanism. If your institution mandates a particular language or cloud, raise it on the call and you'll get a straight answer before anybody writes a scope.

"Don't build this" is one of the four answers, and you get it in twenty minutes rather than month four.

Bring a specialist and we'll scope it live, on your problem. For the institute above, the four came out like this.

Buildable: Their papers and internal documents as one graph, on infrastructure they already pay for.
First: The parser. A bad extraction poisons everything above it.
Hardest: Getting the specialists to agree when two things are the same thing.
Refused: A chat interface, until retrieval was worth talking to.

Refused has been the whole answer before now. Nobody enjoyed the last five minutes of that call, and it was still the right twenty minutes.