PixelRaise, sovereign AI
FR Book a demo

Sovereign AI for public bodies

Your teams save time. Your data stays in Europe.

An assistant that answers from your procedures and regulations, citing its sources. A target three-week pilot, subject to scoping, using the French Mistral model with a controlled hosting perimeter.

Calculate your ROI Watch the video demo2:22 See our references

Start with a scoped pilot and measure the results.

Illustrative example

An example question

“What does our security framework require for passwords?”

Your framework requires at least 12 characters, mixing upper and lower case letters, numbers and special characters. Multi-factor authentication is mandatory for any remote access, under the version in force since March 2026.

Validate retrieval, citations and refusals on your own test set.

Hosted in the EU or on your server

Clients

Already in production in public bodies

No slideware: teams that use our sovereign AI every day, and talk about it.

  • Territoire Innovant award · AI

    City of Annemasse & Annemasse Agglo: "MIA"

    In production for more than 12 months

    Sovereign AI built on open source components (Mistral Small 3.2), hosted in Europe.

    Awarded the "Territoire Innovant" label, AI category, by Les Interconnectés (June 2026, Paris)

    • 340 active users
    • 17,880 messages processed
    • 1 single GPU

    Official figures from June 2026: adoption keeps growing.

    "With MIA, we are showing that a town can master its digital strategy without depending on outside players."

    Gabriel Doublet, Mayor of Annemasse

    Read the detailed case study

  • County-wide scale

    Calvados County Council

    In production for more than 6 months

    The same sovereign foundation deployed across a whole county (~700,000 residents): an AI assistant connected to the county SSO, European hosting, open source models.

    • 73 active users
    • 12,940 messages processed
    • 1 single GPU

    Proof of scalability

    From 90,000 to 700,000 residents: same architecture, same fixed cost, same sovereignty.

    Read the detailed case study

  • Measure your own pilot

    Results depend on your documents, use cases and adoption. Define the measurement period, active users, errors, access controls and review time.

    Agree conditions and deliverables

The PixelRaise team

Technical credibility forged by expertise

PixelRaise does not sell licences: we deploy and maintain a sovereign infrastructure for you. Based in the Annemasse area, with teams in France and Switzerland.

Method: three weeks to decide on facts

From scoping to the go/no-go milestone, with no commitment to roll out.

  1. W1

    Scoping & installation

    Audit of the use cases, SSO connection, sovereign infrastructure in place, first document bases imported.

  2. W2–3

    Pilot with 20 users: go/no-go milestone

    Real usage, adoption and time saved measured. You decide on the numbers.

  3. Go

    Roll-out

    Opened to all users, up and running in 30 minutes, in-house champions trained.

Advanced RAG pipeline

From your documents to the sourced answer

Extraction (PDF, Office, OCR), vectorisation, re-ranking and inference: all of it locally, with citations of the passages used. Plus NotebookLM-style generations: summaries, FAQs, briefings, quizzes, mind maps, depending on enabled components and the approved processing perimeter. Being rolled out as options: audio podcasts and transcription of your meetings.

Your sovereign infrastructure · European Union Hosted in the EU or on your server
  1. Your files common formats
  2. Extraction PDF · Office · OCR
  3. Chunking into chunks
  4. Vectors vector store
  5. Re-ranking BGE reranker
  6. LLM local inference
  7. Answer sourced + cited

  1. 1 · Extraction

    Multi-format

    Native or scanned PDFs (OCR), recent Office files (.docx, .pptx, .xlsx), HTML, text and source code. Every document in these formats becomes searchable.

  2. 2 · Chunking and vectorisation

    Into chunks, embeddings, vector store.

  3. 3 · Re-ranking

    BGE reranker: sourced and relevant answers.

  4. 4 · Cited answer

    Local inference, with citations of the passages used.

  • Structural frugality

    AI during working hours

    Off-hours shutdown can reduce billed computation if the offer supports it. Check startup, minimum billing, erasure and availability; other services may remain on.

    Illustrative GPU shutdown scenario · Schedule-dependent

  • Fixed cost

    Cost depends on sizing, billed hours and support. Shared infrastructure does not guarantee the same price for every workload.

  • Live monitoring

    Prometheus monitoring of the GPUs and the inference engine, an email report at every start and shutdown, with an alert on failure. Backups defined during scoping.

Meetings · optional, being rolled out

From meeting minutes to the cited answer

Committee, management meeting, working group: the recording becomes a text attributed to each participant, then a first draft of the minutes that your staff validate. Once validated, the minutes join your knowledge base, where the assistant finds and cites them like your other documents.

  • What happens to the voice

    The audio is processed in the European Union, with no third-party service, and kept for as long as you decide. Participants are informed before recording.

  • A first draft, validated by staff

    Each contribution is attributed: you name the voices and correct the text. The minutes follow your template, and nothing is shared without validation.

  • Found, then cited

    Validated minutes become sources like any other: a question about a past meeting gets an answer that points to the right document.

The problem

Unmanaged AI creates data risks

Unmanaged use can expose data or bypass document permissions. Define authorised tools, uses, controls and training.

  • Access

    Access control

    Application, document storage and GPU processing use the agreed hosting perimeter. The tunnel encrypts traffic; administration access, logs and backups still require controls.

    Security & compliance

  • Compliance

    GDPR & AI Act support

    An architecture designed around the regulation: data in the EU, structural minimisation, DPO documentation provided, experimental pseudonymisation, in testing. Your DPO must validate compliance for the actual processing.

    AI & GDPR

  • Exit

    Technical independence

    Components and models are selected for your needs. Verify each deployed version, licence and exit conditions; the entire stack is not presumed to be open source.

    Our commitments

Technology stack

Document components and licences

Components and models are selected for your needs. Verify each deployed version, licence and exit conditions; the entire stack is not presumed to be open source.

The stack, deployed on your side

Components can be replaced subject to their interfaces and licences. Inventory external flows, extensions and administration access for each installation.

YOUR SOVEREIGN INFRASTRUCTURE · EUROPEAN UNION Users · SSO sign-in Collaborative AI portal: OpenWebUI secure chat · knowledge bases · sharing by group RAG + Re-ranking multi-format extraction · embeddings · BGE reranker vLLM · inference engine open source LLM: Mistral Small 24B Sovereign GPU: EU datacentre · renewable energy Across the whole stack: monitoring (Prometheus) · OpenAI-compatible API · backups defined during scoping
OpenAI-compatible endpoints: integrate with your business applications. Connectors, access and outbound flows must be validated for each deployment.

  • Open source LLM

    Mistral Small 3.2 24B Instruct in production, on a single GPU: sized to the actual need, no pointless gigantism.

  • vLLM Engine

    High-performance inference, capacity to measure on your actual workload.

  • Collaborative AI portal

    Secure chat, knowledge bases, notes, sharing by group, SSO sign-in.

  • RAG + Re-ranking

    Multi-format extraction, embeddings, BGE reranker: sourced and relevant answers.

  • Observability

    Configure monitoring, alerts and backups. Service hours and fallback capacity are defined in the contract, with tested recovery objectives.

  • API first

    OpenAI-compatible endpoints to plug into your existing business applications.

Your role

Prepare your decision

Free tool · prioritisation

Where do you start?

Map the automation potential of your departments. Set the intensity department by department and get a prioritisation matrix, quick wins vs long projects, plus the top 5 automations to launch, with an estimate of the hours freed up.

  • Identify your quick wins department by department
  • Prioritise impact vs implementation effort
  • Move straight on to costing the ROI

Start the free audit

QUICK WINS PROJECTS Effort → Impact → 1 2 3 4 5
Preview; the real matrix fills in from your answers.

Built for critical sectors

Designed for organisations that cannot afford a data leak.

ROI simulator

Resources

Understanding sovereign AI

Clear guides for your technical teams and your decision-makers: from model technology to regulatory compliance.

  1. Digital sovereignty

    The three pillars of sovereign AI: data location, compliance, control over the model.

  2. Sovereign LLM models

    Open source vs open weight, an agnostic stack, model comparison (Mistral Small 4…).

  3. AI & GDPR

    Legal bases, DPIA, and the right to erasure inside a RAG system.

  4. EU AI Act

    Risk levels, obligations for deployers, applicable deadlines.

  5. AI API pseudonymisation

    Protecting sensitive data even with ChatGPT, Claude or Gemini. Experimental, in testing.

  6. ROI calculation method

    How we put a figure on the data risk avoided (annual loss expectancy).

Contact

Test sovereign AI on your own documents

A 30-minute demonstration on a sample of your real documents, then a three-week pilot (go/no-go milestone included). An expert replies within 48h.

Book a demo · 30 min

Pick a slot directly, or fill in the form.

Email
info@pixelraise.com
Address
251 Route des Marais, 74250 Fillinges, France

Fields marked with an asterisk (*) are required.

One click sends your request: a PixelRaise expert gets back to you within 48 working hours.

Information is used to handle your request and protect the service. Privacy policy

PixelRaise demo2:22 · no sound

Simplified interface, fictitious documents and data. In order, the video shows:

  1. a question asked in plain language on a base of council resolutions, and an answer that cites its sources;
  2. the sources window: the exact extract, the document and the relevance score;
  3. knowledge bases shared by department, and the import of PDFs (including scanned ones) and recent Word, Excel and PowerPoint files (.docx, .xlsx, .pptx);
  4. specialist assistants, including an 80-page tender file summarised in one page with references to the articles;
  5. an HR assistant that answers from the working time rules;
  6. optionally, once rolled out, an audio version of a briefing;
  7. sovereignty: sign-in with your usual credentials (SSO), processing in the European Union, the open-weight Mistral model, installation and maintenance by PixelRaise;
  8. a recap of the offer, including two options being rolled out: meeting transcription and audio versions of files.

Download the video (MP4, 8 MB)