MaraponeAI
One engine, five programs, from the tender to the buyout.
In partnership with Marapone. Built with the Marapone team — they brought the domain and the operational data, we brought the model and systems work.
Visit marapone.com- What we did
- Model training, systems architecture, hosted platform build — with Marapone
- Timeframe
- 2025 — ongoing
- Discipline
- AI
- Status
- Released · 2025

MaraponeAI is the engine the rest of Marapone is built on: a model fine-tuned on the documents that actually run a construction project — drawing sets, project manuals, tenders, subcontractor quotes, RFIs — rather than a general-purpose model asked politely to pretend.
It is Marapone's own model on Marapone's own infrastructure. Nothing a customer uploads is forwarded to OpenAI, Anthropic or Google, and nothing they upload trains anything. That is the part of the original promise that survived; the part that didn't is below.
- 5
- Programs on one login
- 0
- Third-party AI vendors in the path
- $49
- CAD/month to start
01
The problem
Every AI vendor selling into construction arrives with the same general-purpose model behind the same chat box, and construction documents defeat it. A clause in a project manual, a cell in a door schedule and a note on a sheet are three different shapes for the same fact, and the expensive mistakes live in the gaps between them.
The other half of the problem is commercial. The documents a contractor puts through a tool like this — live tenders, subcontractor pricing, their own cost structure — are the most competitively sensitive things they hold, and handing them to an AI vendor whose next customer is the firm bidding against them is not a trade most people want to make.
02
What we built
A private inference stack running models fine-tuned on real construction documents, wrapped in a domain layer that knows a CSI section number from a sheet number, and hosted on infrastructure Marapone operates end to end. No commercial AI API sits anywhere in the path.
On top of it sit five programs — Blueprint Auditor, SpecChecker, AI Estimator, Bid Leveler and ScopeGuard — that hand each other their numbers rather than making an estimator retype them. Each customer's documents are processed in a workspace scoped to their account alone, nothing is used for training, and any project can be exported or deleted on demand.
03
We changed the model, and said so
Marapone used to sell these programs outright: one price, install it, we walk away. It was a clean promise and the wrong shape for the product. Building codes get amended, rate tables go stale, and the model reading a drawing set is measurably better than it was six months ago — a copy sitting on a laptop from last spring gets none of that. Half of what shipped started decaying the day it was delivered.
So the whole thing moved: Marapone runs the engine, the programs open in a browser, and customers subscribe from $49 a month with every improvement arriving the day it ships. It is a harder business — the renewal has to be earned every month rather than banked once — and it means being straight about the trade-off, which is that documents now come to Marapone to be read. Firms whose contracts genuinely forbid that get a private deployment into their own tenant as an enterprise arrangement, negotiated up front rather than implied by a marketing line.
Built with
Fine-tuned open-weight models, Python, Web app, Stripe, Hosted infrastructure
Built on the platform
Two industry suites, one engine.
01/The Construction Suite
Drawings in, awarded subcontract out, nothing retyped.
A bid moves through five distinct readings of the same pile of documents, and each one is a job somebody does badly at 11pm because there isn't time to do it properly. Measure the drawings. Check the spec book against them. Price it. Level the sub quotes that come back. Review the scope before you award.
The suite is one program per stage, in that order: Blueprint Auditor takes the quantities off the sheets, SpecChecker finds where the manual and the drawings contradict each other, the AI Estimator turns quantities into a defensible number, Bid Leveler normalises a dozen quotes onto one scope, and ScopeGuard catches what the winning sub excluded. Each stage hands the next one its output, so the takeoff becomes the estimate and the levelled position becomes the subcontract scope.
All five open in a browser tab on any machine — nothing to install, nothing to licence — and the jurisdiction they price and check against is a data pack rather than a hardcoded assumption. Toronto and Ontario are the deepest, because that is the ground the team came off.
marapone.com/construction- 5
- Programs, one continuous job
- 4
- Plans, $49 to $549 CAD/month
- ∞
- Unlimited use — no credits, no overage
Inside the suite
The companion app
On site, in the pocket
The programs live where the documents do, but half the work happens on site — hands full, no desk. The mobile app closes that gap: scan a plan, check an audit finding, pull up the estimate from the phone already in the superintendent's pocket.
It is the same account and the same workspace as the browser, included in every plan with nothing extra to buy — a subscription covers both at once, and there is no second mobile bill. Native on iOS and Android.
iOS + Android · Q4 2026, included in every plan
The tools, in full
Built with
Private LLM, Python, Web app, Swift, iOS / iPadOS, Android, Computer vision
More work from the lab
Taking one new project for Q4 2026
Have something worth building?
A game, an app, an AI model, or a problem that doesn't have a name yet — tell us about it and we'll build the prototype.