The BLT data engine

Every build teaches the next one.

BLT AI didn’t learn construction from the internet. It learned it from receipts photographed at supply house counters, hours clocked in at 7 a.m., roof decks shot before the drywall went up, and budget lines that went over — millions of small interactions from real jobs, each one a fact about what building actually costs.

Start for free How it works
Receipts read1.2MLine items pulled off paper and posted to a phase
Photos analyzed840KFrames checked for safety and quality issues
Hours logged3.4MCrew time tied to a phase and a location
Budget events6.1MEvery change, approval, and actual, with a who and a why

Cumulative across all BLT projects. Figures update monthly.

01 · The stream

Nothing happens on a BLT job without becoming an event.

A receipt photographed, an hour clocked, a purchase request approved, a phase closed, a budget line moved and the reason someone typed for moving it. Each one is written to an append-only stream with a timestamp, an actor, a project, and a phase. Nothing is a snapshot; everything is a sequence.

Event stream · Maple Ridge #1 · Friday
06:04recapMorning brief generated — two safety flags still open from Thursday
07:02time.inCrew of four clocked in at 7405 Nicholson Rd, located4
07:14vision.flagFall protection not visible on main roof surface — routed to PMsafety
09:31pr.approvePurchase request for Ferguson plumbing rough-in approved1,240.00
14:41ocr.extractHome Depot #4821 — five line items assigned to Framing · Materials550.17
15:08budget.moveFraming current raised, reason: “lumber up since baseline, per ABC quote”+3,000.00
15:09signal.riskFraming materials projected at 96% once pending receipts clear96%
16:22time.outCrew out, 6.5 hrs each posted to Framing · roof deck26.0

Illustrative stream. Real events carry the same shape.

02 · The engine

Six parts. Each one turns something physical into a number.

BLT AI is not one model answering questions. It is a set of readers and comparers that run continuously against the stream, each with a narrow job and a clear output.

Extractpaper → ledger

Reads receipts, invoices, and statements down to the line item. Quantities, units, SKUs, tax. Assigns each line to a project, a phase, and a category, and keeps the photograph alongside so a human can check the machine.

Inreceipt photo, PDF invoice, bank export
Outcoded line items, vendor, date, total
Visionphoto → flag

Checks every frame the crew uploads against safety and quality patterns. Fall protection, house wrap below grade, exposed conductors, standing water, debris on a deck. A flag carries the reason in plain language, never a score.

Insite photo, phase, location, hour
Outflag with a reason and an owner
Baselinehistory → expectation

Holds what each phase has actually cost across every build that closed — by market, by plan, by square foot, by crew. This is the memory that makes “38% above comparable” a fact instead of a feeling.

Inclosed phases, actuals, hours, scope
Outexpected cost and range per phase
Signalstream → risk

Runs the live stream against Baseline continuously. Pace, committed exposure, variance against the estimate a line was raised under. It speaks when a number stops making sense, not on a schedule.

Inlive events, committed money, baseline
Outa claim, its arithmetic, one action
Recapyesterday → standup

Reads everything that happened since you closed the laptop and writes the brief you’d have written if you had time. What moved, what’s waiting on you, and what needs a photo before the inspector arrives.

Inovernight events across every project
Outa morning brief and an attention queue
Askquestion → answer

Plain-language queries against your own transactions. “What did we spend on plumbing?” “Average cost per square foot for foundation work?” It answers from the ledger and shows the rows, so you can check it.

Ina question, in words
Outa number and the rows behind it
03 · The loop

The estimate gets better because the last build finished.

Most construction software forgets a job the moment it closes. BLT keeps it, because a closed phase with real actuals against a real scope is the most valuable thing a builder produces and almost nobody keeps it in a form a machine can read.

Step 01

Capture

The crew photographs, clocks in, and raises requests. Ten seconds each, on the phone already in their hand.

Step 02

Read

Extract and Vision turn paper and photos into coded lines and flagged frames, with the source kept alongside.

Step 03

Compare

Signal runs the live job against Baseline — this phase, this market, this plan, against everything that closed before.

Step 04

Act

A claim with its arithmetic and one thing to do. The decision stays with the human; the arithmetic doesn’t.

Step 05

Close

The phase locks with real actuals against real scope, and becomes the expectation the next build is measured against.

Step 05 feeds step 03. Every closed phase sharpens the comparison for every job that follows — yours first, and the pattern across the platform second.
04 · What the data teaches

Three things no spreadsheet remembers.

05 · Your data

Your numbers stay yours.

Your projects, your costs, and your vendor pricing are visible to your organization and nobody else. Cross-project comparison runs inside your own account first — this lot against your last three, not against a competitor’s.

Where BLT learns across the platform, it learns shape rather than substance: that a phase like this tends to run long, that a photo like this usually means a missing detail. Never your line items, your margins, or who you buy from.

Export anytimeCSV and media archive30 days after cancellation

Scoped to your organization

Every event carries an org, a project, and an actor. Nothing crosses that boundary without you moving it.

Yours to take

Expenses, budget history, photos, and reports export as CSV and a media archive. We don’t hold your data hostage.

Add one build to the record.

Free on one project, full AI. Your first closed phase is the one that starts making the next estimate honest.