Your Data. Your Infrastructure.
Our Intelligence.
Enterprise-grade security by design. ARGUS was built from day one for regulated industries where data sovereignty isn't optional — it's mandatory.
Your Data Never Leaves Your Perimeter
The deployment goes where your documents already are: your own servers, your private cloud, or an air-gapped machine with no outbound path. Where buying hardware is not worth it, inference runs on our own GPUs inside the EEA — which is a different arrangement from sending documents to a provider outside it.
- Full on-premise deployment option
- Our own GPUs in the EEA when you do not want to buy hardware
- Air-gapped installation where the network has no outbound path
- No document content sent to third-party providers
Your Choice of AI Provider
Not locked into OpenAI? Neither are we. ARGUS works with any LLM provider — or run fully open-source models on your own hardware.
Switch providers without changing your integration. Zero vendor lock-in.
No-Training Guarantee
"Your data is NEVER used to train any AI model — not ours, not third-party, not ever."
This isn't just a policy — it's architecturally enforced. Your documents are indexed locally, queries are processed in your environment, and model weights remain completely isolated from your data.
Where a hosted service differs by construction
Not a scoreboard against named competitors — those tables are written by whoever publishes them. What follows is what can be checked on our side, and what follows from the nature of a hosted service, whoever runs it. A good SaaS may well handle the middle rows properly; the first and the last are structural.
| Question to ask | With us | With a hosted service |
|---|---|---|
| Deployment inside your own network | Possible: models run on your hardware | Not by design — the service is the vendor's infrastructure |
| Where inference happens | Our own hardware in the EEA, or your servers | Wherever the provider operates, often outside the EEA |
| Your documents used for model training | No — written into the contract | Depends on the plan and the terms in force at the time |
| Deletion on request | Documents and derived indexes removed together | Usually supported; the derived artefacts are the part to ask about |
| Choice of model | Open-weight models; a different one can be swapped in | The provider's models, changed on the provider's schedule |
| Air-gapped installation | Possible where the network has no outbound path | Not applicable |
Questions asked first
Are you ISO 27001 or SOC 2 certified?+
No. We say it here rather than showing a badge and sorting it out later — a certificate claimed and not held is the one thing a buyer in a regulated industry checks first. What we can do is help with your own certification: mapping controls to your documents and producing the evidence each one needs.
Where exactly do our documents go?+
Either nowhere — the deployment runs on your own servers — or onto our hardware inside the EEA. There is no third option in which content is passed to a provider outside it. The processing locations are named in the contract, not described in general terms.
Are our documents used to train models?+
No, and this is contractual rather than a policy statement on a page. We run open-weight models we did not train and do not fine-tune on customer material; there is no pipeline that would carry your documents into training even by accident.
What happens when we ask for deletion?+
The documents and everything derived from them go together — the index, the embeddings, the cached passages. Derived artefacts are the part worth asking every vendor about: deleting a file while its vector representation stays is deletion in name only.
Who on your side can see our material?+
During a project, the named people working on it, for the duration of that work. There is no standing access for support staff and no shared account. Access is logged, and the log is available to you rather than described to you.
Can it run with no outbound network at all?+
Yes, on your own hardware with no path out. That removes the class of question this page is about, and it costs more: the hardware is yours, updates become a scheduled operation, and the model catalogue is whatever you install. For some institutions that trade is obviously right.
How do we verify any of this rather than believe it?+
Ask for the processing locations and the subprocessor list in writing, ask what happens to derived artefacts on deletion, and ask to see an access log from a past project. Those three answers separate a data-protection posture from a paragraph on a website — including ours.
Ready for a Security Deep-Dive?
Schedule a call with our security team to discuss your specific compliance requirements and deployment options.