Sgaifuture

Practice

Practice

Everything listed here is done by the studio, at one pace, by the people who took the brief. There is no second bench waiting in another city to pick up the dull parts.

Choosing

We turn down about half of what arrives.

A useful first filter is whether the work lives in a body of text you already own. If the valuable knowledge is in people’s heads and has never been written down, an assistant will only sound confident. We ask for a sample of the real files in the first week, under a confidentiality agreement, and we read them.

The second filter is ownership. Someone in the organisation has to care about the answers enough to sit with us on the evaluation set. Without that person, the tool becomes a demo that nobody trusts on a busy morning. We would rather stop after discovery than spend your money on a build that will sit unused.

When a request is really a content mill, a public chatbot, or a wish to score people, we decline. The rest of this page is the work we do take, written so you can see the shape before we talk.

Services

Six pieces of work, each with an ending you can hold.

Discovery week

Five days inside the way you already work. We sit with the people who touch the documents, read a slice of the archive, and watch where a morning actually goes. The week is long enough to see whether retrieval, a pipeline, or an evaluation set would help. We come back with a short paper: three options, a rough shape of effort, and a plain view on whether building anything is worth it. Sometimes the honest answer is to tidy the files and stop.

What you get at the end

A paper of a few pages, the three options, and a go or stop that we will stand behind on a call.

typically 1 week

Assistant over your own documents

A search and answer tool that stays on your contracts, protocols and internal rules. Each reply carries a pointer to the source so a person can open the page and check. Access follows the folders you already share. We keep a log of queries so you can see what is being asked, and so the evaluation set can grow from real use. The assistant does not draft new policy or speak for the organisation in public.

What you get at the end

A working assistant, source citations, access rules that match your folders, and a query log you can read.

4–8 weeks

Document and intake pipelines

Incoming PDFs, scans and mail often land as a heap that a person then types into a sheet. We extract the fields you already track, assign a type, and write a row into the table or CRM you use now. Doubtful cases — a scan that will not read, a message that matches no type — go to a short queue for a person. That queue is the point of the design, because exceptions are where a clerk’s week actually goes.

What you get at the end

A running intake path, a human queue, and a written map of the types and fields the path understands.

3–6 weeks

Evaluation harness

Before we change a prompt or a model, we want a number that means something to the people who will live with the answers. We gather one hundred to three hundred questions taken from real work, write an agreed answer with you, and score the model against that set. The same set is re-run when a file is added, a prompt is edited, or a vendor updates a model. Failures are readable: which question slipped, which source was missed.

What you get at the end

The question set, the scoring script, and a simple way to run it again after a change.

2–4 weeks

Hand-over and training

We sit for half a day with the people who will use the tool and the person who will keep the files in order. We walk through the ordinary path, the failure path, and the two or three knobs that are safe to turn. You leave with a written guide, access to the repository, and a list of the things that should wait for us. The session uses your documents, on your desks, with the questions that come up that morning.

What you get at the end

A live session, a written guide, repository access, and a short list of what to leave alone.

typically 1 week

Quiet retainer

Once a month we look at the same three things: whether a vendor model has moved, whether the evaluation scores have drifted, and whether token spend has jumped for a reason you would want to know. Small fixes sit inside that visit. Larger changes go back to a scoped piece of work. The retainer is optional. Some organisations run the harness themselves after hand-over. Others want a named day in the calendar. We keep the visit short and write down what we saw.

What you get at the end

A monthly note, the scores, a line on spend, and the small repairs that fit the visit.

typically 1 day a month

Tools

On what we build.

We work with hosted programming interfaces when the documents can leave the building under your rules, and with models that can sit next to the files when they cannot. The choice is not a matter of taste. It follows from where the documents live, who is allowed to see them, and how painful a round trip to another region would be.

We prefer the simplest arrangement that still respects those constraints. A small model next to the archive, with a careful retrieval step, often does the job that a large remote model is being asked to do. When a hosted interface is the right call, we still keep the evaluation set and the logs on your side so you can see what left the building.

We do not lock the work to a single vendor for the sake of a partnership. If a model is withdrawn or a price jumps, the harness tells us whether a replacement still answers the hundred questions. That is the whole of the strategy.

Limits

Work we will not take.

We do not generate streams of public copy. We do not fabricate comments, ratings or testimonials. We do not build tools that score people for hiring, credit, or discipline. We do not put a model in a position to make a medical or legal decision without a named person who still has to sign the thing.

Those limits are practical as well as ethical. The studio is small. The evaluation work we know how to do assumes there is a document to check against and a person who will read the doubtful cases. When that structure is missing, we are the wrong workshop.

If a brief sits on the edge of these lines, we say so in the first conversation. It is cheaper for everyone than discovering it in week four.

Next

If the list above sounds like your week.

Send a few sentences about the documents and the part of the day that stretches. We will tell you, within a working day, whether it looks like a discovery week or like someone else’s kind of job.

Start a conversation Read the method