Your tender drawings, your owner's emails, your fee schedules — none of it should ever be on a third party's training set or leaked to a competing GC. The whole point of a Marapone build is that the model and the data live where you put them. Here's exactly how that works.
Security & Trust Pack →
Data-flow diagram, deployment models, and data-handling defaults in one printable PDF for your IT team.
Architecture →
How the system is built and where each component runs.
Integrations →
What it connects to, and what stays inside your walls.
Security & disclosure →
Security overview and vulnerability disclosure contact.
Marapone is based in Toronto and Rome. We do the build work from those offices. The system itself runs wherever you want it: on your laptop, on a server in your office, or in your cloud tenant.
CANADA
Toronto residency
Default for GCs anywhere in Canada or the US. PIPEDA-aligned defaults.
EU
Rome / EU residency
Default for European clients. GDPR-resident processing.
ENTERPRISE
Your own cloud tenant
Deployment into your AWS or Azure tenant, under your keys and your retention policy. Enterprise tier only, quoted as a project.
Take none of this on trust. Each one below is a property of how the system is built, not a policy we promise to follow, which means each one has a check you can run yourself, on your own documents, inside the free trial and before the money is committed — seven days on Solo, fourteen on Field.
One thing to be clear about before you read them: Marapone is a hosted service. Your documents are uploaded to infrastructure we operate, in the browser and in the MaraponeAI mobile app alike, and there is no configuration that keeps processing on your own machines. Everything below is about what happens to those documents once they are with us. It is not a claim that they never left.
RULED OUT 01
The system is built with no client for OpenAI, Anthropic, Google, or any hosted model provider. Inference runs on our own open-weights model, on infrastructure we operate, so there is no code path that could send a drawing or an RFI to one of them.
Check it yourself
We will show your team the architecture diagram and name the exact service that handles inference, under NDA, before you commit to anything.
RULED OUT 02
Uploads are processed to serve you and never used to train a model — including ours. There is no fine-tuning step that touches a customer's documents, on our infrastructure or anyone else's.
Check it yourself
We will walk your team through how the model was built and on what, under NDA, before you commit to anything.
RULED OUT 03
Your documents are read in a workspace scoped to your account. No other subscriber's account reaches it, no project of yours is visible from another firm's session, and nothing is merged into a shared corpus. On Firm, each project gets its own workspace again inside your account.
Check it yourself
Ask the programs about a project you have not uploaded to that workspace, in any wording you like. There is nothing there to find.
RULED OUT 04
Nothing you upload is used to train a model, ours included, and there is no third-party AI vendor in the path to train anyone else's. The model does not learn from your tender package, so nothing from one GC's work has a route into another's answer.
Check it yourself
Ask any client we refer you to whether their build has ever received an update they did not install themselves.
If your procurement team asks for certifications, here is the straight answer before you spend a call on it.
Not held today
We will pursue a third-party architecture review, a penetration test of the mobile service before it leaves beta, and SOC 2 Type I if enterprise volume justifies the cost. We would rather say that plainly than imply a certification we do not have. If your procurement process requires SOC 2 for anything you upload to a vendor, we do not currently clear that bar, and you should know it before the call.
What you get instead
NDAs standard
Mutual NDA on the first call if you'd like one. Project-specific NDAs available before any build starts.
Audit trail
Every model query, every document touched, every response is logged with a timestamp and the user. Searchable. Exportable.
Retention policy
You set the retention windows per document class — RFIs 7 years, daily logs 10, owner emails by your own policy. We don't.
Deletion on demand
A single command (or UI button) wipes a project's documents, embeddings, and any project-specific fine-tuned weights from the system.
Tender drawings are sensitive, and owners and architects often have explicit clauses against sharing them with a third-party service. Read yours before you subscribe: Marapone is a third-party service, your drawings are uploaded to us, and no wording on this page changes that. What we can tell you is that they go no further, no sub-processor sees them, no AI vendor is in the path, and nothing trains on them. If the clause is absolute, do not buy a plan; ask us about deployment into your own cloud tenant instead.
We can provide a Marapone-side NDA template that explicitly covers tender drawings, project schedules, owner correspondence, and trade pricing. Ask for it during the discovery call.
Each project gets its own namespace in the vector store. A user with access to Project A cannot retrieve content from Project B unless explicitly granted. This is a default, not a paid tier.
Everything above describes software that runs where you put it. The MaraponeAI mobile app, shipping Q4 2026, is the one product where that is not true: a phone cannot hold a 48,000-line engine and a model, so the five engines run on infrastructure we operate and the phone is the screen. A document you audit in the app is uploaded to be audited. There is no version of this page where that sentence is softer, so it is the first one.
What you give up
What holds anyway
No plan changes the data path.
Field, Pro and Firm differ in seats, programs and priority. Where your documents are read does not change. Paying more does not make the processing local, and paying less does not make it less private. The reading happens on our infrastructure on every plan, governed by everything in this section.
There are two arrangements. Every plan is the hosted one. The second exists only at the enterprise tier, for firms whose contracts will not allow the first, and it is a project, priced as a project, and it is not a checkbox on the pricing page.
| Plan — hosted by us | Enterprise — your own cloud tenant | |
|---|---|---|
| Where documents sit | Our infrastructure (CA or EU) | Your tenant, your region |
| Encryption at rest | Managed keys, our infra | AWS KMS / Azure Disk, your keys |
| Network exposure | Public API over TLS | Private VPC |
| Access control | Account login, per-account workspace | SSO + IAM + role |
| Third-party AI vendors | None — our own model | None — our own model |
| Training on your work | Never | Never |
| Retention | Yours to set; export or delete on demand | Your own policy, applied in your tenant |
| Air gap | Not available | Not available — the tenant is yours, but it is still a network service |
We used to sell a laptop install and an on-prem server, and both columns are gone from this table because both are gone from the business. If a truly disconnected install is a hard requirement on your project, we are not the vendor for that job and would rather say so here.
Mail security@marapone.com. It reaches the people who wrote the code. We confirm receipt, and we tell you what we did about it.
The machine-readable version of this is published at /.well-known/security.txt, per RFC 9116, so your scanner finds it without a human in the loop.
There is no paid bounty. We will credit you by name if you want that, and not if you do not.
We secure the service. The following stay on your side:
We have a one-page security overview, a data-flow diagram and a draft NDA ready to send.