SA Strategic Agent Grounded on your own records
An application of AI Integrator

It answers from your own records —
or it tells you it doesn't know

Strategic Agent reads the paper you already have. Scans, old files nothing opens any more, spreadsheets, photographs of tickets taken on a phone. It turns them into readable copies you can search, and when you ask it a question it answers by quoting the line it came from and naming the document. When your records don't carry the answer, it says that instead of making one up.

Grounded, or silent

56,687Files converted in the proving run
3Lines, maximum, per answer
2Citations, maximum
0Of your records used to train anything

Most tools that read your documents will answer anything you ask. They sound the same whether the answer is in your files or invented. This one is built the other way round: it can only say what your documents say, and it stops when they stop.

The two halves

read · answer
Half one

Your filing cabinet, made readable

Twenty years of paper becomes twenty years of readable, searchable documents — each one carrying a note of where it came from and how it was read. Nothing is thrown away and nothing is silently skipped.

How conversion works ▸
Half two

An answer you can check

Ask a plain question. Get back the actual line from the actual document, with the document named and the line numbers given. You can open the file and see it for yourself.

How answering works ▸
The rule over both

And a refusal you can trust

"Your records don't contain enough to answer that with confidence." The most important sentence the product says. It is not a failure state — it is the reason the other answers are worth having.

When it abstains ▸

What makes it different

four things

It stays on your machine

The files stay where they are. The model that reads your text runs on the machine itself. Nothing has to be sent to anyone, and it refuses to start a component that would.

What local means ▸

It does not train on you

It looks things up and quotes them. Your records never become part of a model — not ours, not anybody's. That is a property of how it is built, not a promise on a page.

What we do not do ▸

It does not do the maths

Language models are unreliable at multi-step arithmetic. So the model retrieves and quotes, and ordinary code does the sums — the kind you can read, test and check.

The arithmetic ▸

Nothing becomes true without you

The software proposes, and shows where it got it. A person accepts or holds. Only accepted things enter the record. That is what makes a number on a bid worth standing behind.

The human gate ▸

Built on twenty years of real paper

This was not proven on a tidy demonstration set. It was built and run against the owner's own business records across seven industries — drilling, energy construction, environmental, agriculture, equestrian, real estate and biotech. Scans, field tickets, legacy files, whole drives of it. 56,687 files went through conversion; 53,776 came out as readable documents, 98% clean, with two set aside for a person to look at.

That corpus is the founder's own. It is the proving ground, not a customer list — there are no customers yet, and this site will say so until there are.

Who it is for

four shapes

Contractors

Price the next job from your own last ten, with the comparables cited. The estimate, the tickets and the compliance file in one place.

Open ▸

Farms

The shoebox, read. Input costs, feed, herd and land records, searchable and cited when the accountant or the programme asks.

Open ▸

Engineers

A practice's calculations, reports and specifications — the working record of how you decide things, findable years later.

Open ▸

A household

Insurance, the lease, warranties, tax paperwork, a decade of scans. Ask it a plain question about your own life.

Open ▸

What is not finished

Photographs of documents are not searchable yet. Some very old files have no reader. The queue of scans waiting to be read by machine has not been worked through. And the refusal — the thing this product is built around — has been measured once and was not good enough: it answered questions it should have declined. That measurement, and what is being done about it, is published in full rather than summarised.

Read the record ▸