SATORISatori Datum™ · for constructionPrivate AI for the project record
Get the most done.
SATORI reads every document on your project and answers from them, with the line attached. It tells you when the answer isn't there. It runs on a private SATORI appliance, off every AI company's servers.
The latest takeoff I have reads 1,423 MV terms, which 3M has grouped into 321 termination kits. That total comes from the Rev 5 preliminary takeoff, not a final as-built count. The older Rev 3 shows 253 kits, so the kit count moved from 253 to 321 between revisions.
3M flagged it as preliminary: "Kit selections are based off information provided - One-lines. We still require cable specs." I would not sign a term count from the takeoff alone.
The problem
Every job builds a library nobody can read.
The answer to most field questions is already written down. It is in a spec, an RFI response, a submittal comment or a PO. Finding it is the job.
What one project looks like. One electrical scope on a single hyperscale data-center campus: 4,970 files, 209 million characters of text. At about 3,000 characters to a printed page, that is roughly 70,000 pages. Five jobs like it is 350,000 pages. No one reads that. SATORI does, every night.
What the job's own registers show. 299 electrical submittals in ten months. 91 RFIs on the contractor's log, and 223 across all trades. On the latest dated RFIs, a median of 10 days to get an answer, the same as the industry's 9.7.
Why it works in construction
It knows what a spec, an RFI and a PO are to each other.
General AI search treats your documents as a pile of words. SATORI was built independently by an electrical QA/QC manager, proven on a live job, and it follows the rules inspectors follow.
Spec to submittal to RFI to PO
Documents are linked on CSI section, RFI number, sheet and equipment tag. 632,757 references on one project.
The current revision governs
Revisions are grouped into families and the newest is read first, even when your words only match the old one.
It finds conflicts
When a spec says one support spacing and a checklist says another, both are quoted side by side with dates.
Counts from the ledger
"How many" questions are answered from the full ledger, counted by the system. Not a guess from a sample.
Scans and site photos
CARL reads scanned submittals and describes site photos with his own vision model, so they join the record.
Field language and typos
A word the project has never seen is matched to the nearest word it has. Questions from a phone in a truck still land.
How it answers
CARL doesn't search your documents. He knows how they connect.
Before anyone asks a question, SATORI has read every file on the job and knows how they relate: spec to submittal, submittal to RFI, RFI to PO, on to the checklist and the photos. So when a question comes in, the right documents are already one step away. How SATORI does this is proprietary.
It already knows the job
Every spec section, RFI, submittal, sheet, PO and equipment tag on the live project is connected before you ask. It keeps up with your folders every 2 minutes.
The right documents, first time
Asked about MV cable end caps, it pulled the governing spec, both end-cap submittals, the reference sheet, the signed PO and the termination checklist, in under half a second.
Construction's rules, built in
The newest revision is read first. The project's own spec governs over a sister project's. "How many" questions get the whole ledger. Known disagreements come along automatically.
It knows what it hasn't read
It will not tell you the project lacks something until the whole record has been searched. When the answer isn't there, it says so.
Every quote checked
CARL writes a plain answer, and every quotation is checked against the file it came from before you see it. Failures are shown, never hidden.
Sharper every week
Every miss your team reports is traced, fixed and pinned with a test, so the next answer on your job is better than the last.
Unlike anything else in the field
Other tools do one of these. SATORI does all of them.
Each column is a whole category of product, not one company. Most AI in construction today is someone else's cloud reading a slice of your files. SATORI is a private AI reading all of them, on a dedicated box, and proving what it says.
| Capability | SATORI + CARL | General AI chat GPT, Claude, Grok, Gemini | Enterprise AI search Copilot-style assistants | Construction platform AI cloud CDE add-ons |
|---|---|---|---|---|
| Runs on a private, dedicated appliance. Documents never go to a public cloud AI | ✓ | ✗ | ✗ | ✗ |
| No AI company's servers ever process the documents (OpenAI, Anthropic, Google, Microsoft, Amazon, Meta or xAI) | ✓ | ✗ | ✗ | varies |
| No per-seat, per-question or credit fees for the AI | ✓ | ✗ | ✗ | ✗ |
| Reads the whole project folder and keeps up with it live | ✓ every 2 min | ✗ what you upload | connected sources | files inside the platform |
| Links spec → submittal → RFI → PO → checklist → photo on CSI section, RFI number and equipment tag | ✓ computed | ✗ | ✗ | varies |
| Project memory with no size limit: the record lives on its own machine and each question pulls what it needs | ✓ 372,660 passages | ✗ context window | index, no checks | varies |
| Reads the newest revision first, by rule | ✓ | ✗ | ✗ | varies |
| Checks every quotation against the file and shows the ones that fail | ✓ 98.9% | ✗ | links, not checked quotes | varies |
| Searches the whole record before it says "not in the documents" | ✓ | ✗ | ✗ | varies |
| Finds disagreements between documents before anyone asks | ✓ 77 on one job | ✗ | ✗ | varies |
| Counts every row of a ledger instead of estimating | ✓ | ✗ | ✗ | varies |
| Reads scans and site photos with its own vision model, on site | ✓ | in the cloud | varies | varies |
| Gets better on your project: every miss becomes a permanent test | ✓ | ✗ | ✗ | ✗ |
| Exclusive in your market: your competitors can't buy it | ✓ partner right | ✗ | ✗ | ✗ |
"Varies" means some products in the category advertise it and others do not. Based on public product descriptions and pricing pages as of September 2026.
Benchmarks
Measured on a live job, not a demo.
Every number below comes from the reference installation: one active electrical scope, 9 to 28 September 2026, on the reference appliance.
Accuracy
A quote that fails verification stays in the answer, marked ✗. That is the 1%.
Speed and scale
| Task | Measured |
|---|---|
| Find the documents for a question | 0.3–0.4 s |
| Full answer, thinking on | 2–8 min |
| Quick answer | 54–79 s |
| Check the folder for changes | 0.12 s |
| Rebuild 4,970 files from cache | 194 s |
| Describe 34 site photos | 221 s |
| First day with the field team | 26 questions |
Carl vs the big models
Same questions. Same documents. Same grader.
CARL comes in three sizes of the same engine. CARL Compact runs on the reference install today. CARL Pro is the next size up. CARL Max is the full-size engine. The scores below were published by the engine's developer and by an independent lab.
Full size: CARL Max vs Claude and GPT
| Benchmark | CARL Max | Claude Fable 5 | Claude Opus 4.8 | GPT-5.6 Sol (max) |
|---|---|---|---|---|
| PaperBench (research tasks) | 93.0 | 88.8 | 80.3 | 90.5 |
| IFBench (following instructions) | 82.8 | 63.5 | 62.2 | 72.7 |
| WideSearch (finding information) | 81.9 | 81.2 | 72.9 | — |
| HealthBench | 60.2 | — | 52.4 | 55.3 |
| PRBench Finance | 58.3 | 55.8 | 51.9 | 55.5 |
| PRBench Legal | 57.6 | 57.6 | 52.7 | 57.6 |
| CoWorkBench (office work) | 74.8 | 75.9 | 72.3 | 71.5 |
| WorkSpaceBench | 67.7 | 68.7 | 66.8 | 65.6 |
| MRCR v2 256K (long documents) | 92.9 | — | 83.2 | 93.8 |
| LongBench v2 (long documents) | 66.3 | — | 69.1 | 67.1 |
| GPQA Diamond | 92.6 | 92.6 | 92.0 | 94.1 |
| Terminal Bench 2.1 | 86.6 | 84.6 | 84.6 | 88.8 |
| SWE-bench Pro | 67.7 | 80.0 | 69.2 | 64.6 |
| Humanity's Last Exam | 43.6 | 53.3 | 45.7 | 47.2 |
Green = best in the row. CARL Max leads 6 of 14 rows, including following instructions, research work, finance, legal and health, and trails on coding and the hardest exam questions. Scores published by the engine's developer, August 2026 (vendor-run). Their table did not include Grok or Gemini; see the independent index.
Running today: CARL Compact vs Claude Opus 4.6 Max
| Benchmark | CARL Compact | Opus 4.6 Max |
|---|---|---|
| SWE-bench Pro (coding) | 61.7 | 53.4 |
| IFBench (following instructions) | 79.5 | 62.5 |
| OmniDocBench 1.5 (reading documents) | 91.1 | 86.6 |
| CharXiv (reading charts) | 83.7 | 66.0 |
| RealWorldQA (photos of the real world) | 85.9 | 73.9 |
| OSWorld-Verified (operating a computer) | 84.3 | 72.7 |
| AndroidWorld (operating a phone) | 81.9 | 62.0 |
| MathVision | 90.0 | 65.5 |
| LiveCodeBench v6 | 90.3 | 88.8 |
| SWE-MM | 38.6 | 27.1 |
| GPQA Diamond (graduate science) | 89.2 | 91.3 |
| Terminal Bench 2.1 | 73.0 | 78.2 |
| NL2Repo-Bench | 42.3 | 47.6 |
| Humanity's Last Exam | 30.8 | 40.0 |
| CoWorkBench (developer in-house) | 70.7 | 68.2 |
| Developer coding test (in-house) | 79.0 | 63.8 |
The smallest CARL wins 12 of 16 rows (10 of 14 leaving out two of the developer's in-house tests), including reading documents, charts and photos, and following instructions. Developer-published, August 2026.
Independent: Artificial Analysis Intelligence Index
Artificial Analysis Intelligence Index v4.3.2, fetched 28 September 2026, highest reasoning setting for each model. CARL sizes are the published scores of the engine at each size. CARL Compact ranks first of 142 open models in its size class.
What the numbers mean for you
- At full size, CARL sits with the big names. CARL Max scores 45 on the independent index: one point behind Grok 4.7, ahead of GPT-5.6 Terra and Gemini 3.1 Pro Preview. Only the newest Claude and GPT-6 flagships score higher.
- Even the smallest CARL beats a Claude flagship at the work that matters here: reading documents, charts and photos, and following instructions exactly.
- It is the same engine at every size. Moving up from Compact to Pro to Max is an appliance upgrade, not a new product. The vault, the checks and the answers stay the same.
- Compression costs almost nothing. The 4-bit build CARL Compact runs has 0.3% higher perplexity than full precision and picks the same next word about 95% of the time (independent tests).
- On your project, the test that matters is below. Whether the right page reaches the model, and whether its quotes are checked, decides the answer. That part is SATORI.
On the live job: 407 real field questions, by type
Every question the field team asked from 9 to 28 September 2026. Each quotation in each answer was checked against its source file.
| Question type | Questions | Quotes verified | Rate |
|---|---|---|---|
| General project questions | 112 | 1,496 / 1,515 | 98.7% |
| Status | 79 | 2,900 / 2,926 | 99.1% |
| Spec requirements | 59 | 534 / 534 | 100.0% |
| Counts and totals | 42 | 307 / 315 | 97.5% |
| Submittals and POs | 33 | 599 / 600 | 99.8% |
| RFIs | 28 | 456 / 471 | 96.8% |
| Finding documents | 15 | 111 / 112 | 99.1% |
| People and responsibility | 14 | 279 / 281 | 99.3% |
| Schedule and dates | 11 | 161 / 162 | 99.4% |
| Conflicts between documents | 8 | 66 / 67 | 98.5% |
| Lists | 6 | 34 / 34 | 100.0% |
| All questions | 407 | 6,943 / 7,017 | 98.9% |
Graded against the documents on the same job: the 12-question field evaluation scored 12 of 12 with 277 of 277 quotes verified, and the 36-question wide evaluation verified 1,154 of 1,163 quotes (99.2%).
Scales with hardware
Same CARL. Bigger appliance. Bigger brain.
CARL is one engine in three sizes. SATORI specifies, supplies and runs the appliance CARL runs on, and upgrades it as you grow. Your company funds the appliance and pays for the managed service. Nothing in the software changes between sizes. CARL's memory, the whole connected project record, lives on its own machine and can grow without limit: each question pulls in just the passages it needs.
CARL Compact
- Index score
- 34
- Full answer
- 2–8 min
- Answers per 8 h
- ≈ 288
CARL Pro
- Index score
- 40
- Answers
- faster, many at once
- Serves
- whole project teams
CARL Max
- Index score
- 45
- Class
- big-model tier
- Documents
- still private
What does not need the AI engine
Keeping the record current is light work: 4,970 files rebuilt in 194 seconds, changes checked in 0.12 seconds. The AI engine does the heavy reading and writing. Adding projects costs storage, not a bigger engine.
The brain can be upgraded
CARL's engine runs on SATORI's own appliance. Moving from Compact to Pro to Max means upgrading the box, not changing the product. Any new model runs beside the current one and replaces him only if it wins on your own project's questions. On 18 September a lighter build tied on quality but lost on speed, so it was not swapped in.
Compact numbers are measured on the live job. Index scores are the engine's published scores at each size (Artificial Analysis Intelligence Index v4.3.2).
The money
The time and money are already being spent. SATORI gets them back.
Four ways SATORI pays, from the office down to the foreman: hours not spent hunting for documents, rework not caused by the wrong document, RFI effort not wasted, and AI seat fees never paid. Pick a preset, then put in your own numbers. Every line shows its arithmetic.
Each preset opens on the industry median: that is the base case, the floor. The sliders only move up from there, toward the best case your own numbers support.
A model with your inputs, not a measured result. The industry figures are cited on each line. Your pilot replaces the assumptions with your own measured numbers.
The edge
What it does to the contractor bidding against you.
Foremen get it too
CARL isn't only for the office. Foremen and field leads ask from their phones, on six 10-hour days, and get the same checked answer the PE would.
Your people have the answer on hand
Every foreman and inspector gets the memory of your best project engineer: every clause, submittal comment, RFI answer and PO, in minutes, with the line attached.
You catch it before you install it
Old takeoffs, superseded bulletins, and specs that disagree with checklists come up before the pull, not after the failed inspection. One job had 77 of these on record.
Fewer RFIs, faster answers
Industry median wait for an RFI answer is 9.7 days. When the answer is already in the record, SATORI finds it in minutes and the RFI never goes out.
You can take the private work
Owners who won't allow their documents in a public cloud AI can still hire an AI-equipped contractor. SATORI runs on a private appliance.
It gets sharper on your work
Every question your team asks, and every miss it fixes, makes the next answer better. The next bid, the next kickoff and the next new hire start from what the last job proved.
No token meter
Cloud assistants bill by the token and by the seat. SATORI runs on its own appliance, so asking one more question costs nothing extra.
Private by design
A private AI that no public cloud ever sees.
Cloud AI rents you a brain by the token and reads your documents on a public provider's servers. SATORI runs its own model on a dedicated appliance, placed at your site or hosted privately by SATORI.
| SATORI | General cloud AI chat | |
|---|---|---|
| Where your documents go | To a private SATORI appliance only. Never to any AI company's servers. | Uploaded to the provider's servers. |
| Sees the whole project | All 4,970 files, linked, live. | What you paste or upload into a chat. |
| Knows which revision governs | Yes, by rule. | No. |
| Checks its own quotes | Every quote, against the file. Failures shown. | No. |
| Says when it isn't in the documents | Yes, after searching the whole vault. | Usually answers anyway. |
| Cost per question | No token meter. | Metered per token, per seat, per month. |
| Stays current | Folder checked every 2 minutes. | Re-upload by hand. |
| Keeps working without internet | On the local network, yes. | No. |
We have not yet run CARL head to head against cloud models on the same questions. Client documents never go to a cloud model, so that test will use non-client documents and be published with receipts.
Stays current by itself
Plugs into the systems your team already uses.
MILLS watches your project sources and folds every new or changed document into the map. Nobody uploads anything. Documents flow into the private appliance; nothing flows out to an AI company.
Connected today
- OneDrive and SharePoint project libraries, synced live
- Network file shares
- Outlook email saved into the project folder
- Every file type on a job: PDFs, scans, Excel, Word, photos
- Checked every 2 minutes; new files readable the same day
Can be added, with your IT team's clearance
- Autodesk Construction Cloud (ACC): RFIs, submittals, sheets
- Procore: RFIs, submittals, daily logs, drawings
- CxAlloy: commissioning checklists and issues
- Trimble: project documents and logs
- Outlook mailboxes directly
- Monday.com: boards and tracking
None of these is connected today. Each can be added through the vendor's official API once your IT team clears it, and is built to order during onboarding.
The NDA problem
Every cloud AI sends your owner's drawings to an AI company's servers. SATORI sends them to none.
Copilot, ChatGPT, Claude, Gemini, Grok, Meta AI and Amazon's assistants all do the same thing: your question and the documents behind it are processed on the AI company's servers. On a data-center job it is worse, because Microsoft, Google, Amazon and Meta build hyperscale data centers themselves. A hyperscale owner's one-lines, capacity data and equipment schedules can end up on AI servers run by that owner's direct competitor.
SATORI keeps them off all of those servers. CARL runs on SATORI's own appliance, and the only way in is a private, encrypted tunnel behind an email sign-in.
What's in a data-center set
Power capacity, redundancy, equipment, layouts and schedules: the things a hyperscaler's competitors would most like to know.
NDAs limit who sees it
Owner confidentiality terms commonly limit disclosure to named parties or require consent. Adding an AI provider, especially one that competes with the owner, may not be allowed.
"We don't train on your data"
Cloud AI vendors promise not to train on customer data. That covers what the vendor does with it. It is not your owner's permission to share it.
The industry already worries
42% of AEC technology leaders rank data-sharing security as their top AI challenge. Construction firms citing privacy as a barrier rose from 22% to 30% in a year (RICS 2026).
Off every AI company's servers
No document, passage or question ever goes to OpenAI, Anthropic, Google, Microsoft, Amazon, Meta or xAI. CARL runs on SATORI's own appliance: at your site, or hosted by SATORI under your confidentiality terms. SATORI builds no data centers and sells no cloud.
The only way in is a private tunnel
The appliance opens no ports to the internet. People reach it through an encrypted Cloudflare tunnel that connects outward only, after signing in with their work email and a one-time code. The server rejects anything that did not come through that sign-in. SATORI's own machines sync over end-to-end encrypted private links.
An answer for the owner's security team
When an owner asks how your AI handles their documents, you have a clean answer. Competitors using cloud assistants may not.
This is not legal advice. Whether a given tool is allowed under a given NDA depends on its terms; have counsel read your owner agreements. Company roles as of September 2026.
It gets better
Every question makes the next answer better.
The record grows every night. Every miss your team reports is traced to its cause against the documents and fixed with a permanent test.
Ask
Every answer is saved with its receipts.
Keep the best
Fully verified answers join the gold set: 181 so far.
Trace the miss
Graded against the documents, root cause found.
Fix it for good
New rule, new test. It cannot quietly come back.
Re-run the eval
Nothing ships if it scores worse.
Same questions, before and after
What changed on the job
| Measure | Before | After |
|---|---|---|
| Verified quotes per full answer | 2–15 | 6–74 |
| Real documents linked to nothing | 678 | 21 |
| Gold answers | 16 | 181 |
| Stress test, 26 questions: bugs | 9 found | 9 fixed |
8–21 September 2026. The gold set is also the training data for a project-tuned CARL, which is the next step.
The suite
One record. Several ways to use it.
Reads the folder, connects everything, keeps it current.
Compute Analysis Reasoning Logic. The AI. Answers from your documents with receipts.
The map and the Ask box, from any device.
The connected record of your job, with its own trust reports.
PDF markup for inspectors.
Schedules distilled to what's due and overdue today.
One-line drawings traced source to load.
Manpower against scope. In development.
How it is bought
An exclusive licence, a private appliance, and a service that keeps it sharp.
Private AI of this kind is sold the way the market already buys enterprise software. Below are industry reference points so you can compare. They are not SATORI's prices.
| What the market charges for | Industry norm, 2025–2026 |
|---|---|
| Construction project platform | Procore: annual fee by annual construction volume, unlimited users, no public rate card. Reported around 0.1–0.2% of volume. |
| Construction AI add-ons | Procore AI: flat-rate starter pack for up to 3 projects, then credit-based annual tiers. Bluebeam with AI: $590 per user per year. |
| Per-seat construction tools | Bluebeam $260–590 per user per year. Autodesk Build about $1,400–1,700 per user per year. Togal.AI $299 per user per month. |
| General AI assistants | Claude Team and ChatGPT Business $20–25 per seat per month. Microsoft 365 Copilot $30 per user per month on top of the base license. Glean reported at about $45–65 per user per month with a 100-seat minimum. |
| Perpetual license maintenance | 18–22% of the license fee per year. |
| Fully managed IT services | About $110–185 per user per month. |
| Enterprise pilots | 30 to 90 days, paid, fee often credited to year one. |
| What the category is worth | Construction contract-AI company Document Crunch was bought by Trimble for $246.4M in April 2026 with 400+ customers. Trunk Tools has raised $70M. |
The pilot
Sixty to ninety days on one live project with five to ten users. Success agreed in writing on day one:
- 95% or more of quotes verified
- 18 of your 20 hardest recent questions answered correctly with receipts
- 3 or more logged catches that would have cost field time
- Every reported miss traced and fixed within 2 business days
What the partner gets
- The exclusive right to use SATORI and CARL in your market, before anyone else
- A private SATORI appliance, owned and run by SATORI, sized to your work
- Every answer your team asks, with its receipts
- Your documents never sent to a public cloud AI, and never shared with another client
What's proven and what isn't
No other vendor gives you this page.
Proven on a live job
- provenReads and links every file in a 4,970-file project folder
- proven98.9% of quotes verified across 407 field questions
- provenLive sync, nightly refresh, read-only access by email
- provenFinds disagreements; newest revision governs
- provenUsed by a field team of seven from day one
Not yet
- openAnswers take minutes on today's appliance
- openNo head-to-head against cloud models yet
- openNot fine-tuned on project data yet
- openCAD .dwg files are listed, not read
- openOne reference installation so far
It can be wrong. It can't be wrong silently.
Your competitors are still searching. Your team already has the answer.
Christopher Welker · SATORI · satoridots.ai