Deployment
- Single node, or a cluster as volume grows
- Runs on hardware you already own
- Air-gapped installs supported
ROSA On-prem
The whole agent, inside your building.
When your data is not allowed to leave the premises, ROSA ships as an appliance you own. Channels, the agent, the language model, the index and the database all run inside it.
- 0third parties between you and your data
- 5layers inside the box
- 1appliance, yours to own
One line your data never crosses
The only thing that leaves your network is the reply to your own customer. Everything that reads, decides and remembers runs on the inside.
WhatsApp
Telegram
Web chat
Email
—No model vendor and no third party in between.
Inside your building
- 1Channel gatewayThe only part that talks to the outside.
- 2Bud, the agentYour rules and your prices. Nothing decided off-site.
- 3The language model you chooseRun it locally and nothing leaves the box.
- 4Your knowledge, indexed locallyPrice sheets and catalogues, on your disks.
- 5Conversation databaseA database your own DBA can open.
Four things you own outright
The same product, with the parts that usually belong to a vendor handed to you.
APIs and integration
- Your CRM and ERP, over your own network
- Webhooks and REST, authenticated
- Its own single-tenant panel, installed with it
Operations
- Live health checks and alerts
- Usage per channel and per team
- Backups on your own schedule
Governance
- Approval before a rule goes live
- Audit trail of every decision
- Role-based access for your team
Sized to your volume, not to a price list
Three shapes cover almost every deployment.
Single node
One team, one or two channels.
- Conversationsup to ~5k / month
- Concurrent chatsdozens
- GPUone, chosen at sizing
Cluster
Several teams, all channels, peaks that matter.
- Conversationstens of thousands / month
- Concurrent chatshundreds
- Failovernode level
Air-gapped
No route to the internet at all.
- Channelsinternal only
- Updatesdelivered offline
- Modelsshipped with the build
Figures are starting points for the sizing conversation, not a specification. The hardware is agreed with you before a quote is written.
Open-weight models, served locally
LlamaQwenMistralDeepSeekGemma
Chosen with you at sizing, swappable later. The agent talks to it over a standard API.
From sizing call to your team owning it
- Week 1
Sizing
Your volume decides the hardware.
- Week 2
Install
The stack goes up behind your firewall.
- Week 3
Your knowledge
Your price sheets, indexed locally.
- Week 4
Supervised run
Bud answers, your team approves.
- Then
Handover
The keys are yours.
The three questions security always asks
Data residency
On storage you control, in your own country.
No training on your data
There is no model vendor in the path.
Your own retention
A setting in your database, not someone else’s policy.
Everything that comes with it
Data sovereignty
Nothing crosses the line out of your building.
NVIDIA GPUs
The model runs on hardware you own.
Your channels
WhatsApp, Telegram and email.
Its own panel
Installed on the appliance, for your team only.
Your rules
Prices, limits and escalation stay yours.
Quoted per deployment
Because the right size depends on your volume.
Questions people ask
Only for the channels your customers use. The agent, the model and your data stay inside.
An appliance with NVIDIA GPUs, sized for your volume. We spec it with you before quoting.
It is the same agent, with its own single-tenant panel installed on the appliance. Cloud-only pieces such as Instagram are not part of it.
Per deployment, because the hardware depends on the volume you run. Talk to us and we will scope it.