One brain. Every role.
A deep neural network dedicated to your organization — trained on every capture, every RFI, every bid, every outcome. Not our model with your data in it. Your model.
- Model
- ONE PER ORG
- Parameters
- 1.4B
- Training
- NIGHTLY 02:00
- Answers cited
- 100%
Training loop
How it learns your company.
The Brain is not prompted into existence. It is trained — nightly, on the record your crews create by doing the work, in four steps that run per organization and never leave it.
- 01
Capture and record
Every flight, 360 walk, RFI, submittal, change order, daily log, and pay app lands in one system of record. Nothing is retyped to be filed, and nothing has to be exported to be understood.
INGEST · 21T SQ FT · 340+ SOURCES - 02
Curate
The record is broken down to the objects that carry meaning — a clause, a drawing revision, a capture frame, a cost line — then tied to the job, the trade, and the date it belongs to. Superseded revisions and duplicates are dropped before they can teach the network something that stopped being true.
CURATE · REV-AWARE · DEDUPED - 03
Train
Your network trains nightly on your organization's weights alone. No pooled dataset, no shared gradients, no other company's history sitting inside your model — and none of yours inside theirs.
TRAIN · 02:00 LOCAL · TENANT-ISOLATED - 04
Calibrate per role
Each of the 27 role profiles is scored against outcomes you already know — bids you won, pours that slipped, hazards that turned into incidents — before it is trusted with anyone's morning.
CALIBRATE · 27 ROLES · SCORED ON OUTCOMES
The loop closes — today’s capture is tonight’s training data.
Role projection
The same brain, projected.
A superintendent and a CFO are not asking different systems — they are asking the same network different questions. Pick a role: the projection re-weights, the network lights somewhere else, and the brief becomes the one that role would find waiting at their start of shift.
Before the 06:15 huddle.
Read overnight: four daily logs, last night's dock flight, the week-34 lookahead, and every RFI still open on the tower.
- SEQUENCE
The Level 14 slab pour slips two days if the rebar delivery misses Thursday. Splitting the pour at grid line 7 holds the date with the pump you already have.
CITES: LOOKAHEAD WK34 · LOG 08-12 · GATE CAM 06:02
- CONFLICT
Drywall and overhead MEP are both booked into Level 12 east on Wednesday. Last time these two crews overlapped, rework cost 4 days.
CITES: 360 WALK 08:41 · CREW PLAN REV 6
- OPEN ITEM
RFI-0047 has been open 3.2 days and the answer gates the stair core. The architect closed the last three of these in under a day when you called.
CITES: RFI-0047 · DRAWING A-401 REV C
Before the bid review.
Read overnight: a 214-page tender package, 41 of your past bids in this band, and five years of awarded unit costs.
- WIN PROBABILITY
The Halcyon corridor package scores 0.62. Every one of your last four wins in this band carried a night-work premium; this scope does not yet.
CITES: 41 PAST BIDS · AWARDS 2019–2025
- COST DELTA
Structural steel is carried 6.8% above your last three awarded jobs of this type — $412K on this quantity.
CITES: BID TAB 08 · AWARD BR-2214
- RISK
Clause 9.4 sets liquidated damages at $12,500 per day with no weather relief. Your standard qualification for this is missing from the draft.
CITES: TENDER PKG P.214 · CLAUSE 9.4
Before the capital call.
Read overnight: 60 job cost reports, every open change order, and the installed quantities verified against capture.
- EXPOSURE
$18.4M of unapproved change orders sit across nine active jobs. Four are past their contractual response window and convert to disputes on Friday.
CITES: CO REGISTER · CONTRACT §7.2
- CASH
The June draw is $2.1M light against forecast. Two progress claims are waiting on installed quantities that Tuesday's capture already verified.
CITES: CLAIM 041 · ORTHO 06-11 · 98% VERIFIED
- VARIANCE
Portfolio budget variance is 3.4%, against 11% at the same point last year. Two heavy-civil jobs carry the whole difference.
CITES: 60 JOB COST REPORTS · ERP SYNC 05:30
Before the site walks.
Read overnight: 11,400 captured frames, yesterday's observations, and the certification register for every crew on shift.
- HAZARD
Nine open hazards, ranked. Two are unprotected leading edge on Level 9, detected in last night's dock flight and unassigned since 21:40.
CITES: FRAME 8821 · FLIGHT 06-12 · OSHA 1926.501
- TREND
Housekeeping observations on the podium deck are up three-fold week over week — the same pattern that preceded the March slip incident.
CITES: 47 OBSERVATIONS · INCIDENT 03-19
- CERTIFICATION
Four scaffold tickets expire within ten days for crews already scheduled on the tower.
CITES: CERT REGISTER · CREW PLAN REV 6
Before coordination.
Read overnight: the coordinated model at revision 12 against Tuesday's point cloud, element by element.
- BIM VS BUILT
Level 12 overhead MEP is installed 62 mm below the coordinated model across three bays. The ceiling grid still clears, but the sprinkler drops do not.
CITES: SCAN 06-11 · MODEL REV 12 · TOL ±25MM
- MODEL DRIFT
Seventeen structural elements no longer match the as-built cloud within tolerance. Fourteen trace back to one approved change order.
CITES: CO-118 · POINT CLOUD 06-11
- HANDOVER
The as-built record for Levels 1–8 is 94% complete. The gap is the north stair, uncaptured since the shoring came out.
CITES: 214 WALKS · COVERAGE MAP 06-12
Before the crew steps off.
Read overnight: your crew list, the laydown manifest, yesterday's pour log, and production against the estimate.
- TODAY
Six carpenters on Level 11 layout. Your embeds are staged at grid F-4 in the laydown yard, not at the hoist where the manifest says.
CITES: LAYDOWN SCAN 05:20 · MANIFEST 4471
- BLOCKER
Level 11 east was placed at 19:40 last night. Formwork strip is Thursday, so hold the wall layout on the east bay until then.
CITES: POUR LOG 06-12 · MATURITY 68%
- PRODUCTION
Interior framing is running 340 lf ahead of estimate this week — enough to pull the Level 12 start forward a day.
CITES: 360 WALK 06-12 · ESTIMATE LINE 09-220
Projection weights
- Sequence0.92
- Cost0.34
- Hazard0.61
- Quality0.55
One set of organizational weights. Each role’s head reads a different mix of them — which is why the estimator and the superintendent can disagree without either being wrong.
One network, eight heads
Every AI feature is one head of the same brain.
Progress, Safety, Inspection, Earthworks, Bid, Cost, Legal, and Knowledge are not eight models bolted to eight products. They are heads on one set of organizational weights — which is why a capture anomaly on Level 9 can flag a budget line and a contract clause in the same breath.
Progress AI
Reads every walk and flight and reports percent complete by location and by trade — 80+ trades, no BIM required.
Safety AI
Scans every captured frame for hazards, ranks them by severity, and assigns each one to the person who can clear it.
Inspection AI
Runs standing inspection routes from dock and robot capture, flagging what fell out of tolerance since the last pass.
Earthworks AI
Turns each flight into cut, fill, and stockpile quantities, reconciled against the design surface nightly.
Bid AI
Reads the whole tender — drawings, specs, addenda — and flags buried risk with page-level citations.
Cost AI
Prices work from your own award history and projects cash exposure across the portfolio, job by job.
Legal AI
Reviews contracts against construction law and your own precedent, clause by clause, with the clause quoted back.
Knowledge
Answers from your own record: the RFI you closed in 2019, the detail that failed, the sub who delivered on time.
Governance
An instrument you can audit.
An instrument is trusted because it can be checked. The Brain drafts; your people decide. Every answer carries its source, every inference is logged, and nothing your organization builds ever teaches anyone else's model.
Cited to the source
Every answer names the capture frame, drawing revision, or contract clause it came from. Open the citation and you land on the page, the frame, or the line — not a summary of it.
Full inference audit log
Every question, answer, model version, and cited document is logged and exportable. You can reconstruct what the Brain said in March, and what it read to say it.
Human sign-off gates
Nothing consequential — a claim, an award, a change order, an incident classification — moves without a named person approving it. The Brain drafts; people decide.
No cross-customer training
Your weights are yours. Your data never trains another organization's model, and no other organization's data has ever trained yours.
Aligned to the EU AI Act
Documented data sources, risk classification, and human oversight per deployment. Model cards, data-residency options, and DPAs on request.
Institutional memory
Expertise doesn't retire here.
When a 30-year estimator retires, their instincts are already in the weights — the jobs they priced, the subs they trusted, the margins they protected, the calls that turned out right. The record stays, and the judgement stays with it.
0%
Institutional knowledge lost
27
Role profiles calibrated
1.4B
Parameters per organization
100%
Answers cited to source
“The Brain briefs our board on capital exposure across sixty active jobs, and every number opens to the frame or the clause it came from. Nobody has asked for a slide deck since.”