A state-of-the-art autonomous agentic system, for U.S. plants
First plants at no charge · Applications close 19 September 2026
You get what a data scientist, a planner, an analyst and a finance lead would tell you about your own plant — from an app that runs on your own computers, at no cost, in hours instead of years. The first plants pay nothing. Applications close 19 September; the work starts at the beginning of 2027.
What you get
The work of four specialists, as an app on your own computers
How it works
An app you install. It works inside your building, and nothing is sent anywhere
What it costs
$0. No card, no deposit, no invoice
What you risk
$0. Non-binding, and either side may walk away at any time
Apply now — it costs nothing See what you get Two minutes. Non-binding. Nothing due now. Nothing to install today — the company is in formation and the build starts at the beginning of 2027.
Is this you
Every one of these comes from the same gap: decisions are made faster than they can be measured. Close that gap and all four get easier at once.
A supplier, a run size, a promise to a customer. Each one moves cash. Nobody can say by how much until the year is closed and it is too late to choose differently.
Then: you see the cash effect before you commit, not after.
Margin fell. Was it input prices, freight, and demand, or was it your sourcing and your mix? Without an answer, next quarter is another guess.
Then: the result comes split into the part you caused and the part you absorbed.
The same sheet rebuilt every week. The same reorder judged by feel. The same report retyped for the same meeting.
Then: the small repeating decisions run automatically and you review exceptions.
No structure, no versions, no record of what was decided or why. So a result today cannot be traced back to the choice that caused it.
Then: this is the one to fix first, and it is where the work starts.
The idea
Your margin change is really two changes added together. One part came from outside: input prices, freight, demand, a supplier's own trouble. The other part came from inside: what you sourced, what you ran, what you held, how you staffed it.
Reported together, they cancel out into a single figure you can only shrug at. Pulled apart, they give two different instructions. Volatility you cannot control, you hedge and diversify against. Inefficiency you can control, you fix — and the fix is worth a known amount.
Doing this needs structured data, versions, and a record of what was decided and when. That is why the first work is on your data, not on a slide.
The vision
Safer than a spreadsheet. An Excel file gets emailed, copied and left on a laptop; this is built so your numbers never leave the building at all.
Finance, supply chain, data science and innovation expertise, on state-of-the-art world modelling — not a chat window bolted onto your spreadsheets.
Chatbots talk. This is designed to do the work: financial and economic modelling that holds up in production, that understands what it is doing, and that shows its workings every time.
America is the greatest country in the world. AMR hopes to make it greater still.
Made by an Italian who loves America — and who came here to build.
A statement of intent, not a description of something finished. None of it is built yet — which is exactly why it is written down here, where you can hold it against what arrives.
What you get
Everything below is built to keep working after the Service ends, because a tool you cannot run yourself is just a report with a longer shelf life.
01 · The first Service
When a supplier, a product mix, or an inventory level is on the table, you get the cash effect of each option and a recommendation you can defend in a room.
Layout of the summary page, drawn empty on purpose. No sample figures, because there is no client work to show yet.
02
Sourcing, product mix, and supplier concentration are usually decided on unit price and hope. The Service prices those options from your own numbers, including what concentration is costing you in risk terms, so you can choose before you are locked in.
03
The weekly sheet and the meeting pack still run themselves. Each figure is traced back to the decision that produced it, period over period, so the report is an explanation rather than a reprint.
04
Purchasing, production and payment timing treated as one free-cash-flow policy rather than three separate habits. The first Service delivers the policy design, the objective function, and the data required to train it — not a trained model on day one.
05
The model, the templates, and the tracking stay with you and work the same way next year. No data-science hire, no dependency on me to run it again.
How it works
By 19 September 2026
Plant details and a contact, two minutes. Non-binding, and either side can withdraw at any time.
From October 2026
Every application gets read after the window closes. Confirmations to selected businesses go out from October 2026. If the numbers cannot support the work, I tell you that instead of selling you a study.
Beginning of 2027
The first version is built for your plant, on your numbers, with the savings written down so they can be checked afterwards.
What it costs
A market comparison, not a former price of ours. Left is what five specialists cost. Right is what you pay.
If you hired five specialists
$3,000
15 hours at $175–250/hr. Scoped projects quote from $5,000.
Your cost in this cohort
$0
No card, no deposit, no invoice.
Five separate people, $175–250/hr. A short booking each is about 3 hours — you do not buy 90 minutes of a data scientist.
15 × $200/hr is about $3,000. Here those same five seats are software running on your own computers, and your cost is $0 — with nothing spent on handoffs, briefing, or reconciling five separate opinions. Five half-days at $250/hr, or a scoped project, from about $5,000.
The launch cohort is a bet I am willing to carry, not a discount off a former price of ours. For the first plants, the Service costs nothing. If it works and you want a stronger version afterwards, we agree in writing. Nothing on this page promises a result, and nothing converts to a paid arrangement automatically.
There is no second step to sign up for. If the first one works, the numbers will make the case better than I could — and you will be the one deciding what a second engagement is worth.
Applications close 19 September 2026. Applying does not reserve a place, applying early confers no priority, and I may choose fewer plants, or none. Confirmations go out from October 2026. Work begins at the beginning of 2027 either way. No card is collected.
Why you can trust it
Most tools ask you to upload your numbers to someone else's computer. This one never asks, because it is built so that it cannot use them. Nine questions, answered in one line each.
Nowhere. It runs on your own computers, where your numbers already sit.
The software is installed inside your building and works there. There is no server of mine holding your data, because there is no server of mine in the picture at all.
No. I never hold the key that unlocks them.
The key is made on your machine and stays on your machine. Without it the files are unreadable noise. I never hold the key, so I could not read your numbers even if I wanted to.
No. Nothing records what you opened, asked, or looked at.
There is no usage log to keep, so there is nothing to leak, nothing to hand over, and nothing here for anyone to read later.
On your own hardware, by default. Unplug the network and it still answers.
If a job would need a model larger than your machine can hold, nothing is sent anywhere unless you switch that option on. If you do, that model is to run in a sealed enclave: memory only, nothing written to disk, no logs, no questions or answers kept, checked cryptographically before the session starts, and wiped when it ends. Either way your numbers never train anything.
Nothing of yours. No copies, no backups, no history of what you asked.
This is not a thirty-day or ninety-day retention policy; there is nothing to expire. It also means there is nothing for me to delete on request, and nothing for anyone to subpoena from me.
The design removes most of the question, because nobody else ever holds your data.
Those rules mostly govern what happens when somebody else holds, moves, or looks at your data. Here there is no transfer, no processor to appoint, and no breach surface on my side. Health information is never disclosed to me, so there is no business associate to sign up; export-controlled drawings are never exported. I will not claim a certificate that does not exist yet: the position is to be confirmed by independent U.S. counsel and security auditors before the first plant uses the software. None of this is legal advice about your own obligations.
You do not have to take my word for it — the code will be published.
Open source means you can hire any cybersecurity firm you like and have them read it line by line, then tell you whether what I claim here is what the software actually does.
Independent U.S. security experts, not me.
Before the software touches a plant's numbers, it is to be verified and certified by outside experts who have no stake in the answer.
In the cloud your data sits on someone else's computer. Here there is no someone else.
Whoever runs a cloud service can be compelled, careless, or breached. Treat this like a notepad on your own desk — safer, really, because a notepad cannot tell anyone who opened it.
The rare part is the combination: local only, end-to-end encrypted, no logs, open source, independently certified. Any one of those is common enough. All five together is what turns confidentiality from a promise into a property of the design.
None of this exists today. It is the standard the software is to be built and independently tested against before any plant's data is involved, and the confidentiality duties are written into the conditions you would be signing.
Who does the work
Stefano Ciccarelli
Data Scientist at a Fortune 10 / FAANG company. New York metro, based in Jersey City.
The problem sits across five disciplines rather than inside one, and the Service puts all five in one place.
Why the number can be trusted
Instead of
The same class of work, at a price a $20M plant will not sign, and the model leaves with them when the work ends.
An accurate record of what happened. It cannot tell you which part of what happened was your own doing.
Can move the crate. Will not tell you whether to change supplier, and leaves you nothing reusable for the next order.
Can build a model. Cannot tell you which decision it should optimise, or what the answer is worth in your accounts.
Why now
Talk of bringing work home is up, and imports still rose. That gap exists because most plants cannot price the decision, so they sign on unit price and hope. That is the gap this closes.
$2.93T
U.S. manufacturing output, roughly a tenth of GDP. Federal figures.
35,000+
Small and mid-sized plants in the NIST-MEP universe.
97%
of U.S. exporters are small firms, and they still hold only about a third of export value.
These are counts of how many plants exist and what they ship. They are public figures about the size of the sector, not a claim about how many will buy this.
Later · not part of this window
Not live, and not part of the first 2027 Service. There is a checkbox on the application if you want it noted. It will only be built if plants use the first Service well.
Before you ask
No. Nothing here is live software for sale right now. The company is in formation and work begins at the beginning of 2027. What you can do today is apply to be among the first plants the Service is built for, at no charge.
No. The application records a non-binding expression of interest. Either side may withdraw at any time, for any reason. No payment is due now, and nothing authorises work or billing before 2027.
By hand, by me, after the window closes on 19 September 2026. I am looking for plants where the question is real and the records are good enough to answer it — not for whoever clicked first. Applying early confers no priority and does not reserve anything. Confirmations go out from October 2026. Everyone who applies gets an answer either way, and if you are chosen you sign the letter of interest on this site to make it official.
It is the opposite. Unstructured data with no version history is the normal starting point and it is the first thing the work addresses. What matters is whether enough was recorded to reconstruct past decisions. After a short call you will get a straight answer about that, including "not yet" if that is the truth.
Because the claim is written before the work, in terms your own books can settle, and the market part is separated from the internal part. If a good quarter came from prices moving in your favour, the report says so rather than taking credit for it.
Two minutes to apply. Non-binding, withdraw whenever you like, nothing due now.