Architecture

Eight programs. One architecture.

Most "AI for construction" pages stop at the marketing layer. This one doesn't. Below is what is actually inside the software: the parsers, the reasoning engine, the database file, the data packs, and the one design rule the whole thing is built around — the model is never allowed to find anything. It runs on our GPUs and arrives in a browser tab, so there is no server for you to rent and no hardware to buy.

Get the Stack Walkthrough → See the Products

What exists

Eight programs, two screens.

Eight finished programs — five that price and win the job, three that hold it after award — plus a ninth screen putting the same interfaces in a site super's pocket. They share a codebase spine on purpose: a document read correctly in one app reads correctly in all of them, and the output of each stage is the input to the next without anything being retyped.

APP 01 · TAKEOFF

Blueprint Auditor

How much is there. The heaviest pipeline of the eight: vector PDF, rasterised sheets, DWG and DXF all land in the same measured quantity model.

PDF · DWG · DXF · scanned sheets · schedules

APP 02 · DOCUMENT AUDIT

SpecChecker

Whether the manual and the drawings agree. A cross-check engine over measurable properties, indexed by CSI division, that reports both statements and both references.

crosscheck/ · attribute_rules.json · spec_patterns.json

APP 03 · PRICING

AI Estimator

What it should be bid at. Assemblies resolve against the rate packs; risk runs as a Monte Carlo over the cost model instead of a flat percentage on the bottom line.

assemblies/ · pricing/ · risk/ · scenarios/ · numpy

APP 04 · LEVELLING

Bid Leveler

Whose quote is actually cheapest. Quotes arriving in a dozen formats are normalised to one scope, and every adjustment prints the basis it was made on.

extraction/ · leveling/ · knowledge/ · openpyxl

APP 05 · BUYOUT

ScopeGuard

Whether anyone actually priced it. Cross-references the tender package against every proposal for gaps, overlaps and wording that will not survive a dispute.

crossref/ · master/ · clarify/ · learning/

APP 06 · SCHEDULE

FloatGuard

Whether it still finishes on time. A real CPM solve — early and late dates, total float, broken logic, with a time-impact writeup per delay event.

cpm/ · SS/FF/SF + lag · tia/

APP 07 · COMPLIANCE

TradeVet

Whether the sub is cleared to be on site. Status is derived from each sub’s own document expiry dates instead of by hand, so it cannot go stale unnoticed.

prequal/ · expiry-derived status · scorecard/

APP 08 · JOB COST

CostLine

Whether it still finishes under budget. Cost-to-complete by CSI division, taken from the more conservative of committed value and burn rate.

wip/ · percentage-of-completion · forecast/

APP 09 · Q4 2026

The mobile app

All eight engines, reached from a phone, included in every plan at no extra cost. How it drives the real bridges without a rewrite is below.

Capacitor · FastAPI gateway · 243 operations

The shared spine

Six layers, no black boxes.

Every program is built the same way, and only the domain layer differs — that is what makes it a suite instead of six products that happen to share a logo.

Information moves one way. MaraponeAI reads the engine’s result, never the documents.

LAYER 01 · INTAKE

Your documents, into your own workspace

Project manuals, drawing sets, subcontractor quotes, scope letters, addenda — PDF, DWG, DXF, XLSX, DOCX. Dropped into the window. They go to your workspace and nowhere else, not to a third party, and not into anyone’s training data.

↓

LAYER 02 · EXTRACTION

What the documents actually say

Text layer first, OCR when there isn't one. Shared across the apps and version-pinned together, because a version skew shows up as a PDF that reads in one app and not another.

↓

LAYER 03 · DOMAIN ENGINE

crosscheck/ · leveling/ · crossref/ · pricing/ · engine/

Deterministic code. This is the layer that produces findings — every conflict, gap, overlap and dollar, each one traced to a numbered page of a real document.

↓

LAYER 04 · REASONING

MaraponeAI

Our own model, reading the structured result the layer above produced. It explains, prioritises and drafts the letter. It does not search the documents.

↓

LAYER 05 · STORAGE

Your workspace, versioned

One workspace per account, versioned on every save, so what the documents said at tender is recoverable later, which on a claim is the whole argument.

↓

LAYER 06 · INTERFACE & OUTPUT

The screens, and what leaves them

Reports out as PDF and Excel, on demand, in a form you can hand to an architect or attach to a claim.

The load-bearing decision

The model never finds anything.

Every finding comes out of the parsers and the deterministic engine, from text that exists on a numbered page of a document you supplied, and every one of them cites where it came from. MaraponeAI reads that finished result and writes about it. It is never asked to go looking.

That constraint is not a technical convenience — it is the reason the output is usable. An RFI goes on the project record with your name on it. If the architect answers "that clause does not say what you claim", you have to be able to point at the page. A model that invented a plausible requirement will have made you look careless, in writing, on a project you have to keep working on for two more years.

Both numbers it compared are on the row — you can check the finding without taking its word for it

MODE 01 · LOCAL

The embedded model

Our own model, running on hardware we operate. No third-party AI vendor sees your drawings, there is no API key anywhere in the path, and one model serves the whole suite.

Our hardware · our model · no third-party AI vendor

MODE 02 · RULES

The deterministic fallback

If the model is unavailable, a deterministic narrative generator reads the same structured result and writes the same briefing without free-form language. The product is fully functional in either mode — the findings, the citations and the numbers are identical, because they never came from the model in the first place.

Same findings · same citations · same exports

Layer breakdown

What each layer guarantees.

DOCUMENT EXTRACTION

Reading the set

· The text layer, where a set has one
· Sheets rendered at the resolution small annotations need
· DXF from any CAD package
· DWG read directly — AutoCAD 2004 and later
· Room boundaries measured off the line work itself

Pinned to one version across the suite. A skew would mean a PDF that reads in one app and not another.

OCR

Scanned and plotted pages

· Recognition tuned for drawing lettering
· Plotted sets with no text layer, read from the page
· Rotated, noisy and photographed sheets
· No OCR service, no API key, no upload

A great many drawing sets are plot-to-PDF images with no text layer. A sheet the app cannot read is a sheet that does not get audited.

REASONING

MaraponeAI

· Our own model, on our own hardware
· No third-party AI vendor in the path
· One model serving the whole suite
· Deterministic RULES mode when absent

The system prompt forbids inventing a value, a section number or a sheet number. The architecture makes it structurally unable to anyway.

STORAGE

Yours, and versioned

· Every project versioned on every save
· A version row on every save
· RFI log and finding feedback kept with the project
· Isolated per account, with no pooling

Your annotations on the contract documents are privileged working product. They are yours, exportable on demand, and deletable by you at any time.

INTERFACE

One UI, two shells

· Runs in the browser — nothing to install
· HTML / JS, no framework, no build step
· No plugin, no licence key, no machine to provision
· The same UI runs in the iOS shell

The Blueprint Auditor interface is a single ~20,000-line document. The mobile app runs it unchanged.

DEPLOYMENT

Nothing to install

· Each engine runs as its own server-side service
· Opens in any current browser, or the mobile app
· Every account works in its own workspace
· An update reaches every seat the day it ships

No installer, no Python, no Docker, no runtime to keep patched. Your IT department is not being asked for a project.

Where it runs

In the browser you already have open.

This used to be the section where a page like this quoted you a GPU server. It still does not apply, but for the opposite reason: the GPUs are ours. The analysis pipeline runs on Marapone hardware and the programs arrive as a web interface, so the machine on the estimator's desk needs nothing beyond a browser and a connection. What that costs you in exchange is set out plainly further down.

FROM $69/MO · THE PLANS

Any machine

Any machine with a browser — Windows, macOS, Linux, five years old, borrowed. Nothing to provision, nothing to configure, no admin rights. The reasoning layer runs on our GPUs, so yours is irrelevant.

$0 hardware · live in minutes · no IT lift

$4,900+ · CUSTOMISED SETUP

Your own files

Same architecture, tuned to your company: your historical documents, your templates, your rate sheet, your house language in the drafted letters. Appears in your account beside the eight standard programs.

Engineering work · your data packs · same engine

$15,000+ · ENTERPRISE

Your own systems

Integration into what you already run — Procore, Sage, Autodesk or your own database, with single sign-on, your own retention rules and a signed DPA. Deployment into your own cloud tenant is available at this level.

Your tenant available · your governance applies

The five plans in full →

Local knowledge is data, not code

The jurisdiction layer.

Rates, tax percentages, statutory holdback, permit fees, code citations and trade knowledge are never compiled into the engine. They load at runtime from JSON files that ship inside the app and stay readable, which is why retargeting a market is an edit, never a rebuild, and why you can do it yourself.

Jurisdiction pack

Tax, labour burden, lien & holdback rules, permit fees, code references, winter productivity, bonding thresholds, local by-laws. Every number carries a citation string that prints beside it in the output.

1 file · 14 sections · marked verify_before_bid

Price catalogues

Materials and crew compositions in separate files, with a city cost index applied on top — material, labour, equipment and subcontract factored independently instead of by one blended multiplier.

95 material items · 24 crews · 24-city index

Override layer

Your own rates and an optional supplier feed layer on top of the baseline at resolution time. Your real buyout numbers win without anyone editing the shipped catalogue.

Overrides > supplier feed > baseline

The scope knowledge base is indexed by CSI MasterFormat, so the taxonomy is portable by construction. Toronto/Ontario ships as the default pack because a finished one is the only honest demonstration of depth. See how regions & rate packs work →

The phone app · Q4 2026

The same engines, on the phone.

The honest constraint first: the analysis pipeline — the parsers, the OCR, the geometry, and a model — cannot run on a phone. Any vendor claiming otherwise is running it on their server and calling it mobile. We run it on a server too, and we say so: everything here is a hosted service, phone and browser alike, and your documents are read on our hardware.

What it is not is a reimplementation. Each program exposes its entire feature set through one AppBridge class — the object its own HTML interface drives, and those classes instantiate perfectly well with no window attached. So the gateway drives the real bridge, and the phone reaches 243 operations across the five products with nothing reimplemented. A fix to an engine reaches the app with no porting step, which is the only way a bid priced on a phone stays identical to the same bid priced in the browser.

The shell

A native app built with Capacitor — iOS and Android, with real camera, filesystem and share access. Photograph a page on site and it enters the same pipeline as a scanned PDF.

Capacitor · iOS & Android · native APIs

Eight engines, six services

The five pre-award products each expose a top-level src package and pin incompatible dependencies, so they cannot share one process. The three post-award engines were built later against one dependency set, so those share a single service. Each call spawns that product's own interpreter running a small adapter, exchanging one JSON command.

Subprocess per call · bounded queue · per-tool limits

One workspace per account

Subscribers are isolated by pointing the engines' home directory at a per-account folder. Each account gets its own projects, award history and learning corpus, with zero changes to the engines themselves.

Per-account isolation · content-addressed uploads

The phone is another window onto the same account, and not a separate product: the app is included in every plan, with nothing extra to buy and no code to redeem. What it does, and when it ships →

Maintenance & upgrades

We run it. You can still leave.

Leaving is one export

Your projects live in a versioned store scoped to your account, and everything in it comes out on demand — takeoffs, levelling sheets, audit reports, in formats that outlive us. Cancelling closes the programs; it does not hold your work hostage.

The model gets swapped for you

Nothing above the reasoning layer is tied to a particular model, because nothing above it depends on the model for a finding. When a better open-weight model ships we swap the file. You get it the day it lands, with no version to buy and no upgrade to schedule.

It gets quieter with use

The programs keep a record of what you asked, what the answer was, and which findings you judged to be noise. That corpus makes the next project quieter, and it stays inside your workspace, never pooled with another firm's and never used to train anything.

No lock-in clause

Month to month, no annual commitment, no implementation fee, and cancellation is a button in your own dashboard, never a phone call — any time it stops doing what this page says.

The security posture is mostly a consequence of the architecture

EGRESS

No third-party AI vendor. Your documents are never forwarded on.

DATA AT REST

Encrypted on our infrastructure, in Canada or the EU. TLS in transit.

TENANCY

One isolated workspace per account. No pooling across firms, no training on your work.

AIR GAP

Not available. Documents are uploaded to us; enterprise can deploy to its own tenant.

The full security posture, including what we can and cannot see →

Want to read the actual code?

We’ll open the repo on a call.

Thirty minutes, screen shared: the cross-check engine, the reasoning layer, the database schema, and one of your own documents run end to end. If your IT lead wants to be the one asking, bring them.

Request the walkthrough → Or try the live demo