Loading...
 
Skip to main content
cyber resilience act digital sovereignty gdpr rag self-hosted ai private ai ai assistant ai agents open source models local llm ai audit ai act

Private AI agents and assistants by Open Source Solutions

Private AI agents and assistants

AI assistants built on your organisation's documents and procedures, deployed on infrastructure you control.

  • Hosted on your infrastructure
  • Sharing on your terms
  • You choose where data lives
  • GDPR, CRA, AI Act tracked
  • Sourced answers (RAG)
  • Tested on your real cases
  • Failures logged and measured
  • Open source models
  • Maintenance and updates
  • Performance tracked over time
  • Regular reviews per contract
  • Training for your teams
Companies with contracts, client files, internal procedures or sensitive information need to protect their data when using AI assistants. With an external AI service, the data you enter is processed under the provider's terms of use, configuration and jurisdiction. Once your data leaves your infrastructure, it is no longer under your control. Providers may change policies, get acquired, or suffer breaches.


AI providers start cheap to lock you in, then raise prices once you depend on them. With your own assistant, you control the budget. Built for the way you actually work, it searches your resources directly, without you having to explain the context again, and saves time on repetitive tasks.

We deploy your assistant on your infrastructure or in a private environment you control. You decide where the data is kept, who can access it and how every action is logged. A simple query does not need the same care as a client file: in specific cases, after study and for budget reasons, routine queries can be routed to an external service. Sensitive content always stays internal. It is a documented policy, not a default setting.
Ranking of ten open source models on a 24-question demo audit: gpt-oss 20B leads with 96% correct answers in 8.4 seconds, ahead of larger models.
Human review card for a model answer: requirement "no invented figure, date or name", blurred answer, AI judge suggestion and human decision.
Terminal output showing the deployment safeguard: execution refused because the model is not qualified.
The four stages of a private AI assistant: audit of uses, evaluation on a test bench, signed qualification, deployment only when the model is qualified, with usage feedback looping back to the audit.
Access rules for AI agents: read broadly, write narrowly, execute explicitly, log everything.
Request triage diagram: sensitive content always goes to the internal model, routine queries can optionally go to an external service after study.
Zabbix dashboard of the AI safeguard: protections checked every minute, model Internet attempts blocked, agent writes logged.
Zabbix dashboard during a model test run: progress, GPU load and temperature, answers passed and failed.

A complete solution, not just a model

In short: a private conversational assistant, built around your documents, your procedures and the way you work, without depending on an external AI provider for every request.

We start with an audit and a list of requirements agreed with you, and use it to build a test bench from your documents and your real questions. At the end of the audit you receive a report with the cases tested, the answers obtained, the gaps found, the limits identified and our recommendations. That report is yours, whether or not you continue the project with us.

Depending on the project, we then install and configure the solution and set up the tools needed to use your document bases, publications, archives and other internal resources. We adjust the prompts, instructions, settings and the tooling around the model to improve the results step by step. We also provide documentation and train your users, as you need.

This field moves fast: a model installed today may be outdated in a few months. We plan for it. We test new models on your own test cases, and we switch when the gain is measured, without rebuilding the solution.

From the audit to daily use

We start by defining your uses, your technical constraints and what success means for you, and turn them into a list of requirements to test. From there we build a test bench from your documents, your questions and your real situations. We compare models, their possible quantisation, their memory and computing needs, and the possible architectures. The results let us recommend a model, hardware and an architecture that fit your uses, your budget and your security constraints.

Once you approve, and depending on what you need, we install and configure the solution, prepare the ingestion of your documents, organise access and set up the tools that make them usable: indexing, search, source citations, archiving, or connections to your internal applications. We support your users and adjust the prompts, rules, sources and tools until the answers are truly usable. This phase also shows where the assistant fails, where questions need to be clearer and what should change.

What comes next depends on your needs: maintenance, updates of the models and components, quality follow-up, monitoring, backups, training, continuous improvement, remote support or a change of model. You build your setup freely, one piece at a time. Each piece can be added to an existing server maintenance contract or covered by a dedicated AI services contract.

Frequently Asked Questions

No. The model and the database stay on your infrastructure, on a dedicated environment we operate for you, or on your own hardware. In specific cases, after study and for budget reasons, you can choose to route routine queries to an external service. Sensitive content never leaves.

Open source models, chosen to match your needs and your hardware. You are never locked into one vendor: the model, like your data, stays exportable.

One to two weeks, depending on how many cases are tested.
You fill in an online scoping form (specifications and requirements), we review it together, then we run the audit on our test and evaluation machines. Every candidate model is put against your real cases, several times over, and we measure what fails as carefully as what works. You then receive the report with our findings.

Your access and ours are logged. Every intervention is tracked and dated in your Client Space, viewable at any time. It's not a promise on trust, it's a written record.

Yes, depending on the chosen architecture.A fully internal deployment can operate without a permanent Internet connection when confidentiality or security requirements demand it. Updates, backups and maintenance access can be managed separately.

Yes.
We favor an architecture that limits dependence on a single model or provider. The model can therefore be changed as your requirements, hardware, costs and available technologies evolve, without rebuilding the entire solution.

It is tightening. GDPR, the Cyber Resilience Act, the AI Act: obligations are arriving in waves and already apply to organisations that did not expect them. We track these texts for our own services and we document what applies to your deployment. Cyber Resilience Act compliance

An ecosystem around your AI assistant


An AI assistant never works alone: it relies on a whole ecosystem that keeps it secure and keeps you independent. We can take care of hosting the whole infrastructure, monitoring it around the clock, and hardening its security against hostile environments. Here are some of the services around your assistant:

Offsite backup
Encrypted backup of your ingested content and your vector database, on private infrastructure in France, with documented restore tests. Learn more
Sovereign WireGuard VPN
Restrict access to your assistant's admin dashboard and API to authorised addresses only, through an encrypted tunnel. Learn more
Password management
A company password vault to protect access to your assistant's backend and your tools. Learn more

Let's talk about your needs

Want your own private company AI? Tell us about your situation. We will tell you which first step fits best and what we can offer.

Assess my AI project

privateai_en