Every organization builds a library
nobody can read.
One construction project alone: 5,897 files, more than 70,000 printed pages.
Private AI for the documents that run your business
Every page you own. One mind.
SATORI reads every document your organization runs on: contracts, specs, policies, tickets, spreadsheets, scans and photos. CARL answers your team's questions from them, with the exact line attached, and never leaves your private appliance.
First deployment: 5,897 files on a live construction job · each dot is a document · drag across the field
Every organization builds a library
Ask CARL
Ask what you'd ask your sharpest expert.
Compute Analysis Reasoning Logic
C
ComputeHis own private engine on a dedicated appliance. He reads every page you give him, with no token meter and nothing sent to an AI company.
A
AnalysisTakes apart documents, spreadsheets, scans and photos. Counts whole ledgers, not samples, and finds where documents disagree.
R
ReasoningThinks it through the way your best expert would: which document governs, which version is newest, what is still open.
L
LogicEvery answer follows from your documents. Every quote is checked against its file, and when the answer isn't there, he says so.
These are real questions from SATORI's first deployment, a live construction job, with CARL's actual answers in full, with owner, contractor, supplier and file names removed. Pick one and watch the record light up.
the job's record
CARL · answered from your documents
It says when it isn't there
It can be wrong.
It can't be wrong silently.
Every quotation in every answer is checked against its source file before you see it. When the documents don't hold the answer, CARL says so.
A chat AI, guessing · illustration
What's the warranty period on the transformers?
"Transformers typically carry a standard 5-year manufacturer warranty."
An illustration of the typical confident guess. It comes from nowhere in your documents.
CARL, on the live job
What's the warranty period on the transformers?
"The transformer-specific warranty period isn't in what I have read. The only period I can stand behind is the 1-year end-to-end Contractor warranty for the Work."
6 of 6 quotes verified. Nothing invented.
Measured on a live job, not a demo
Five weeks. 990 real questions. Every quote checked.
First deployment: one electrical scope on a hyperscale data-center campus, 9 September to 13 October 2026. Built independently, proven in the field.
Money saved in those five weeks
$39K–$79K
330 to 660 hours of hunting handed back to the work. 990 questions × 20 to 40 minutes saved each × $119, the job's own cost per labor hour.
At peak, over a year
$2.34M
120 people from the office to the foreman, six 10-hour days. Search time, rework, RFI effort and AI seats, from the value model.
Modeled from measured use. The question count and labor rate are the job's own; minutes saved per question are industry-reported (Construction Dive, 2024). The yearly figure is the base case of the calculator below.
98.7%
of 14,061 quotes verified word for word against the file, across every real question asked
990
real field questions in five weeks, from the office to the foreman. None removed
13,878
quotes matched their source file word for word, in five weeks of real use
5,897
project files on the job, watched for changes every 2 minutes
0.3 s
to find the right documents for a question
$0
per question. No token meter, no seat fees to an AI company
392 / 400
separate graded test: 400 questions with known answers, scored right or wrong
230 / 230
on that test: perfect on status, dates, spec lookups and conflicts
Field figures are every question asked from 9 September to 13 October 2026, with nothing removed; they measure whether each quotation matches its source file. The graded test is separate: 400 questions written with known answers from the job's documents, 28 to 29 September 2026. First pass 370 of 400; every miss was traced, fixed and re-asked, for 392 of 400.
Private by design
Off every AI company's servers.
Copilot, ChatGPT, Claude, Gemini and Grok all process your clients' confidential documents on the AI company's servers, and those companies may compete with your clients. SATORI sends them nothing.
Zero AI companies
No document, passage or question goes to OpenAI, Anthropic, Google, Microsoft, Amazon, Meta or xAI.
No open doors
The appliance opens no ports to the internet. People reach it only through a private encrypted tunnel, after a work-email sign-in.
An answer for your clients
When a client's security team asks how your AI handles their documents, you have a clean one. Not legal advice; have counsel read your NDAs.
What it's worth
The hours are already being spent. SATORI gets them back.
Worked example from the first deployment, construction. Each preset opens on the industry median, the base case. The sliders only move up from there, toward what your own numbers support.
$2.34M
value returned per year · base case at the floor
Search hours returned
Rework avoided
RFI effort saved
Cloud AI seats not bought
hours a year handed back to the work, full-time people's worth of searching.
A model with cited industry inputs, not a measured result. Your pilot replaces the assumptions with your own numbers.
Nothing else like it
Other tools do one of these. SATORI does all five.
In September 2026 we checked 173 AI products across the US, UK, Europe, Australia and Asia against these five capabilities.
01Knows how your documents connect
Every reference between documents, linked before you ask.
02Proof on every line
Every quotation checked against its file. Failures are shown to you.
03Your industry's own rules
Newest revision wins. Whole ledgers counted. Conflicts surfaced.
04A private engine
Its own model on its own appliance, off every AI company's servers.
05Learns your business
Every miss your team reports becomes a permanent test.
173
AI products checked
5
capabilities checked on every product
0
combined all five. None had a clear yes on more than one.
One engine, three sizes
At full size, CARL sits with the big models.
Independent intelligence index, published scores. The same CARL at every size: moving up is an appliance upgrade, not a new product.
Running today
CARL Compact
34
In the first deployment now. Beats a flagship cloud model at reading documents, charts and photos.
Next size up
CARL Pro
40
Faster answers, many at once, for whole project teams.
Full size
CARL Max
45
In the range of the big cloud models, and still private.
It gets better
Every question makes the next answer better.
Ask
Every answer is saved with its receipts.
Keep the best
Verified answers join the gold set: 399 so far.
Trace the miss
Graded against the documents, root cause found.
Fix it for good
A new test for every miss. It can't quietly come back.
Re-run
Nothing ships if it scores worse.
Go deeper
The full story, with the numbers.
Partnership
Partner with SATORI.
A partner gets a private appliance run and kept sharp by SATORI, and exclusive use of SATORI and CARL in their market and industry while the agreement holds.
The exclusive right
While it holds, no competitor in your market can buy it.
A managed service
A private appliance run, upgraded and kept sharp by SATORI. You fund it; SATORI runs it.
A voice in what's next
Your field's questions decide what gets built next.
Christopher Welker · satoridots.ai
Already on a SATORI team?
Sign in to DOTS.
Use your work email. We'll send a one-time code; no password to remember. Only people on your company's access list can get in.