On-prem AI · commercial construction

Nobody reads the whole set.
Now something does.

Plans, specs, schedules, end to end — and every answer comes back with the sheet and the spec section it came from. It runs on a box in your building, so the documents you are not allowed to upload never have to be. We also work on the reading itself: one process, rebuilt with your team, measured before and after.

$4,000 to $5,000, fixed. One day on site or on a call, one process mapped end to end with real cycle times, and a written finding inside five business days. Credited in full against a build if you start one within 60 days.

Direct to us. No form, no sales sequence.

One process, measured before and after Fixed scope, fixed price Nothing goes out to answer a question
redgorge-01 · document set loaded LOCAL

ASKED

Does anything in the mechanical spec conflict on duct insulation?

CONFLICT — SPEC vs. SPEC

Two paragraphs of the same section call for different values on the same assembly. One of them is going to be installed.

SPEC 23 07 13§2.1.Bp. 412
SPEC 23 07 13§3.5.Ap. 415

Sources on this answer: 2 · unsourced claims: 0

Mock-up on an invented set — the shape of an answer, not a run. The real run, twelve of twelve quotes verbatim, is in 03.

01 The work we go after

Every company has a job that is one person, reading, at 11pm.

Not "inefficiency". A specific hour, with a name attached to it. Four questions we ask in the first meeting, because the answers are usually a name, a duration and a failure, and that is a scoped piece of work.

01

What happens when an addendum lands?

Who opens it, what do they do with it, and how long until the field knows. The base documents still say the old thing, and they will keep saying it until somebody reads both.

02

What was the last thing installed wrong because two documents disagreed?

And what did it cost. Not a productivity number. A change order, a demo-and-rehang, a delay you carried.

03

Where has somebody already tried to fix this?

Every company has a graveyard. What was bought, who championed it, and why it did not take. That history is worth more to us than a wish list, and it usually names a budget line that already existed once.

04

If that job took a quarter of the time, who notices?

And what do they do with the time. If nobody notices, there is nothing to fix and we will tell you so on the day rather than three weeks into a build.

02 What we do, and what it costs

One process. One number. A fixed price and a date it ends.

Every engagement carries an acceptance criterion written as a number, not an adjective, and a measurement taken before we change anything. If we cannot measure it, neither of us can tell afterward whether we did anything, and we would rather say that up front.

01

One-day teardown $4,000 to $5,000 · one day, written finding in 5 business days

One process mapped end to end alongside the people who run it — submittals, change orders, RFIs — instrumented with real cycle times off real packages, a ranked list of what to fix, and what each fix is worth per year. You own the finding whether or not you ever hire us again.

Credited in full against a build started within 60 days. It exists so neither of us has to guess at the scope of the bigger thing.

02

Fixed-scope build $18,000 to $35,000 · 4 to 6 weeks

One process and one metric, both named in the scope document with you, measured before and measured after. Submittal log triage. An addendum reconciled against the base documents. Spec sections pulled into a takeoff-ready schedule. An equipment schedule checked against the one-line and the panel capacity.

Half up front, half on acceptance, and acceptance is the number in the scope document rather than our opinion or yours.

03

Larger programs

Several processes, sequenced, each one carrying its own acceptance number. Priced on scope after a teardown, and never as an open retainer.

What we will not do

No staff augmentation. We are not a body shop. If what you need is a pair of hands inside your team, we are the wrong call and we will say so in the first ten minutes.

No hourly. Hourly has no defined end, and it means every tool we build that makes the work faster cuts our own invoice. That is a bad incentive to hand somebody you are hiring to make something faster.

Nothing we cannot measure. If we cannot count the thing before and count it again after, there is no way for you to know whether the money did anything, and we will not take work we cannot be graded on.

03 The proof — Redgorge Reader

Every answer comes back with the sheet and the paragraph. Check it in ten seconds.

Redgorge Reader reads a construction document set and answers questions about it with the citation attached, so you can open the document and check it before you put your name on it. It is not what we are selling. It is how you can check, before you spend anything, whether we are careful with somebody else's documents.

01

The set

Pittsburg Unified School District — District Admin. Offices HVAC Units Replacement, Bid 24-026. Specs dated March 2024, 333 pages. It read the full mechanical division: all eight Division 23 technical sections, 49 pages of specification.

A public bid document, and a public record. Ask and we will send you the exact PDF we ran, so you can check every quote against it yourself — which is the whole point. Ask a vendor for the set behind their demo and see what comes back.

02

The run

Seventy-five seconds, wall clock, start to finish — on the appliance, in the room. Nothing was uploaded, and nothing about that document left the machine.

No OCR and no vision pass required — this manual had a clean text layer.

03

The result

Twelve of twelve quotes verbatim. Every passage it attributed to the manual appears in the manual, word for word, and every paragraph number it cited was real. It also examined six apparent contradictions and called all six correctly as not conflicts — including duct liner against external duct wrap, which are different assemblies in different sections and are the exact false positive a keyword search reports.

That is the only thing this run tested, and it is the one that has to hold first: whether a citation is worth enough that you stop checking them by hand.

Exactly what exists, as of today.

Everything below is either running today or marked not yet.

Running Cited answers with the crop Ask about a sheet in plain language and get the answer plus the sheet, the spec section and a picture of the exact spot, so you can open the document and check it in about ten seconds. The first question on any sheet takes seven to nine seconds, because nothing about that sheet has been read yet; asking again about the same sheet is closer to three. On a set nobody has opened, every sheet is the first one. Every citation it returns is checked against the labels actually found on that sheet; anything it made up is dropped before you see it, and the number dropped is reported.
Running Motor and feeder sizing off an equipment schedule NEC Article 430 arithmetic. And when the schedule totals cover a larger set of loads than the one-line device actually feeds, it refuses to publish a bus rating at all, says which two scopes disagree, and tells you to reconcile them. It does not average them and it does not pick one.
Running Questions across a whole set, and the honest negative When it cannot answer, it says what it read and what it could not read. On the Pittsburg specs it opened with "Searched all 333 pages." On a set with scanned sheets it counts the pages it could not read rather than pretending it read them. A bare "not found" hides whether it looked at all. Two sheets that disagree are reported as a disagreement, never averaged and never quietly resolved. The five-sheet questions on record took between forty-three seconds and a minute and forty, because one machine answers one thing at a time.
Running RFI drafting When the documents genuinely do not answer something, it names the gap, says where it already looked, and drafts the RFI for you to send. There is no model anywhere in that path -- the draft is assembled from what the reader already cited, so it cannot invent a reference.
Running Spec-against-spec conflict checking Built into the app. On a 38-page spec we wrote with four conflicts planted in it, it found two or three of the four in each whole-document pass and never called a near-miss real; the runs are in build log entry 013. Spec against drawing schedule is where the money is, and it comes after the schedule parser above — that number has to come up first for the answer to be worth anything.
In progress Schedules off drawing sheets Pulling a door, window or equipment schedule off a drawing sheet into a table. Reading a spec is one job; this is a harder one, and it is still being built. It is graded against eight drawing sets we read and labelled by hand, so there is a right answer to check against. On the current build about four in ten rows it returns are on the sheet and about five in ten of the sheet's rows come back (macro-F1 0.441, read with its composition: two of the eight sets have no rows to find, and across the six with real labels the mean is nearer 0.25). Nothing changes in it without passing that test against all eight sets; the best score it ever produced, 0.52, was refused because it broke 32 other documents. The run-by-run chart is in the build log.
Not built Submittal and addendum handling Not built yet. It is exactly the kind of thing we build on an engagement, against your real submittals rather than a demo set — which is the only way it would be worth having.
04 Who we are

Twenty years in construction. A decade of it in the model.

It started in the glass business.

Then five years in BIM, taking our own HVAC and process piping into coordination on a California hospital tower, the Sphere, Allegiant Stadium and the Hard Rock — one trade among many in the model, working it out against everybody else's.

Then five years as CTO of an industrial refrigeration contractor: built the BIM department from nothing and authored the FabCAD database the shop fabricates from. Not a report somebody filed. A thing the fitters use every day.

Ten years of that is ten years of reading these sets by hand — every sheet, every spec section, on sets that ran to hundreds of pages. Redgorge is the tool we kept wishing existed. We built it on hardware we can hand you, because we know exactly which of those documents you are not allowed to upload.

Revit · Navisworks · Revizto · FabCAD MEP · reality capture & point clouds
prefabrication programs · Procore · NetSuite

05 Why it runs in your building

The platforms in your contract went through procurement. An AI service did not.

Your document platforms were named, reviewed and signed off, and everybody knows where the files sit. A service with its own retention and training terms went through none of that — so pasting a set into one is a conversation you would have to go and have. This is one quiet box on your network, so there is nothing to disclose, nothing to review, and nothing to explain to the owner.

The other consultant you could hire for this work would paste your set into a hosted model. That is the difference, and it is the entire reason the box exists.

About the box, plainly. The model occupies 41.1 GB of graphics memory while it is answering. A workstation with an RTX 4090 has 24 and is not enough, and almost nobody has forty spare gigabytes of GPU memory idle in a closet. So the hardware is a requirement, not an upsell: one machine, bought at cost after the contract is signed, currently around $1,500 to $2,000. It is a line item against the engagement. We are not in the hardware business and we do not want to be.

Project NDAs and owner confidentialityHealthcare, government, gaming, data centers. Uploading the set to a third-party service is often a straight breach.
Security and life-safety drawingsCamera layouts, access control, egress, utility routing. These are not documents you paste into a chat box.
Your bid strategy is in those filesPricing, means and methods, where you found the money. Competitive information, sitting in someone else's logs.
One box, not per-seat foreverAdd coordinators, add a project, open a second office. It keeps working when the trailer's connection does not.
Nothing goes out to answer a questionEvery run described on this site was produced by the appliance reading its own storage with its own models. Your documents are not sent anywhere to get an answer. Ask your IT manager what they need to see and we will show them.
06 The first step

What actually happens if you say yes.

The whole thing, spelled out — including where your documents go, what it costs, and what happens if it turns out to be worth nothing.

01

You name one process

The addendum one. The submittal one. The one where somebody re-types a schedule into a spreadsheet. You do not need to know what the fix is; naming the job and the person who owns it is enough to start.

If your contract needs an NDA signed first, send us yours.

02

We walk it with your team, stopwatch in hand

Whoever runs it walks us through a real package — they know where it hurts, we bring the stopwatch. Real cycle times, the actual sequence rather than the one on the org chart. Where documents are involved, we would rather work on real ones, and where they cannot leave your network, we bring the machine to them or we work on your side of the wall.

Tell us what should happen to any files afterward — deleted, returned, or never copied off your network in the first place — and that is what happens. Nothing is used to train anything. Nothing about your company appears on this site; the tests published here are public bid documents, and they stay the only ones.

03

You get it in writing inside five business days

The process mapped end to end the way your team actually runs it, what each step costs you per year, a ranked list of what to fix, and what a fix would take. We go through it with you before it is final, so anything we misread gets corrected. Yours to keep, and yours to hand to somebody else if you would rather they built it.

Including the parts we could not measure and the parts we got wrong. That list is the most useful part for us and usually the most reassuring part for you.

04

Then you decide, with a number in front of you

If a build makes sense, the teardown fee comes off it in full. If it does not, we part on good terms and you still have the map. If the honest answer is that this process is not costing you enough to be worth fixing, that is in the document too, in writing, and it is the cheapest way to find out.

If it earns nothing, you have lost a day and the fee, and you know something about your own operation that you did not know on Monday.

Or, if you would rather grade us first

Send one closed-out job you already know the answers to. Plans and specifications together if you have them. We run it, and you get back what it found and what it got wrong, every claim cited to a sheet. No fee, and no obligation to do anything afterward except tell us the truth about it.

A set you already know is the only fair test, because you can check what it found against what actually happened in the field. Do not tell us what went wrong on it. Keep that back and grade us on it.

Start here

Pick the job nobody wants to do on a Thursday.

One day, one process, one written finding you own. If a build follows, the day comes off the price. If it does not, you keep the map.

Teardown: $4,000 to $5,000, fixed. Credited in full against a build started within 60 days. Grading a closed-out set costs nothing.