Deployment
Each option changes where your data goes, which suppliers process it and who runs each control. We write that down for your deployment before any documents move.
Three ways to deploy
New to the terms? Read what private AI means or how on-premise and self-hosted differ.
Run by us
We host and run it. Storage region, model provider and support terms are agreed with you in writing.
Agreed before you sign
In your account, run jointly
Marella AI can be deployed into a customer-controlled cloud environment after architecture review.
Agreed before you sign
Your servers, including on-premise
A self-hosted deployment package is available; prerequisites and responsibilities are agreed during architecture review.
Agreed before you sign
Data-flow review
Your files are not the only copy. Review the text extracted from them, what is sent to the model, and the logs and backups as well.
Original files, connected records, versions, deletion and access changes
Parsed text, page coordinates, chunks, metadata, embeddings and extracted structures
Queries, retrieved context, prompts, model/provider requests and generated outputs
Identity events, decision traces, security events, logs, support access, backups and recovery copies
Who does what
For each one, write down who sets it up, who runs it and who signs it off: sign-in, storage, models, patching, monitoring, backups, incidents, deletion and support.
The workflow, the kinds of data involved, where data may be held, your approved cloud and model providers, your identity and network policy, and who would operate the deployment.
No single region applies to every customer. Where data is stored depends on the deployment you choose, and we state it in writing during review.
Not by default. Cutting a deployment off from the internet means designing and testing its network, model access, updates, licensing and support route for that.
That depends on the deployment. The order form and responsibility matrix name who owns each, how they get access, the recovery expectations and how issues are escalated.
Yes, but changing where data is stored, how people sign in, which model runs or who operates it means re-testing security, performance and support.
Keep the first message non-confidential. We agree a secure route for diagrams, questionnaires and sample documents afterwards.