Sovereign company AI: BrainDup, the local RAG

Ask Marvia about the history of uranium in the Morvan. Within twenty seconds she answers, cites her sources, and lets you click each one to check. Marvia is a marten, the emblem of the Morvan nature park, in the heart of Burgundy. Above all she is an artificial intelligence that has ingested 1,423 documents about the region. And everything she knows fits in an appliance sitting on a desk. Unplug the internet: she keeps answering.

Marvia is the demonstration of BrainDup, the open source project AS3P revealed in its first showable version on 4 August 2026, after nine months of work. Back in April, in our analysis of NemoClaw, in French, we called for an appliance that works without an external router and without a cloud fallback. BrainDup is that concrete answer.

The dilemma a sovereign company AI resolves

Every company faces the same pressure: adopt AI without exposing sensitive data. Sending your email, your proposals, your technical sheets and your trade secrets to an American or Chinese supplier raises a confidentiality problem. The GDPR fits a foreign public cloud badly. The AI Act adds another layer of requirements. For a smaller firm under strict regulation, the question is often a blocker.

A sovereign company AI reverses the premise. Instead of sending your data to the AI, you bring the AI to you. BrainDup is a RAG system (Retrieval-Augmented Generation): it does not retrain a model, it fills a database that a local language model queries. Filling a database is an everyday act in a company. Training an AI remains the preserve of the technology giants and their compute budgets.

How BrainDup holds up without inventing

BrainDup is deterministic. It refuses to answer when it lacks sufficient context. It would rather say “I do not know” than make something up. That guarantee rests on an architecture of three databases, acting as four successive safety filters.

PostgreSQL holds the truth about provenance, the permissions and an immutable audit trail. No user reaches a document they have no right to.
Milvus, the vector database, finds the relevant passages in a fraction of a millisecond through semantic search, over 1024-dimension vectors.
Neo4j, the graph database, maps the relationships between entities and enriches the context.

The principle is strict: the code decides, the language model executes. Citations are assembled by the code from the catalogue, never by the model. A false citation is treated as a blocking bug. It is one of the 22 architectural invariants that govern the engine.

BrainDup and the AI Act, compliant by construction

The timing works out. On 2 August 2026 the AI Act became applicable. AI systems must now clearly state that they are AI, and the European Commission supervises providers of general-purpose models such as ChatGPT, Gemini or Claude. For sensitive uses, companies must put in place data governance, technical documentation, human oversight and traceability. Penalties can reach 35 million euros or 7 % of worldwide turnover.

A sovereign company AI meets those requirements by construction. Data governance is native: PostgreSQL, roles and permissions. The audit trail is immutable. Human oversight is built in through curator arbitration. The documentation ships with the code. And because the system is local and open source, you depend on no model provider under Commission supervision. Compliance is not a layer added afterwards, it is a property of the architecture. That is also the argument of those who criticise the AI Act: without an industrial strategy, regulation subjects Europe, and open source remains the only lever of sovereignty for mid-sized companies.

The proof through Marvia

A sovereign company AI is worth nothing without its demonstration. For Marvia, AS3P ingested 1,423 documents and around twenty websites about the Morvan. Result: 29,515 text fragments, 30,887 entities detected in the knowledge graph. Every answer carries its sources, with clickable links to the original PDF or to the cached version of the site. Sources can vanish from the internet; they stay in your RAG.

The system rests on six open source pillars proven for local use: Django for the interface, n8n for automation, FastAPI for machine-to-machine exchange, Ollama to run the models, Docling to convert documents, GLM OCR for poor-quality scans. Everything is in the box.

The cost, without the token trap

BrainDup runs locally, with no token cost. A built-in calculator puts a figure on the gap: processing 20,000 documents at a cloud service's rate would come to 9,812 euros, whereas here the investment stops at the hardware, once. Under the Apache 2 licence the code is auditable and yours. You change language model whenever you like, without depending on a vendor.

From user to builder: Dompteur and Chaman d'IA

Adopting BrainDup also means building your own skills. AS3P offers a clear ladder. The Dompteur d'IA course teaches everyday use of generative AI. The new Chaman d'IA course goes further: it makes you the owner and independent operator of your own sovereign AI, through the BrainMad method of AI-assisted development. You leave with a working BrainDup on your own real case. Chaman d'IA is in pre-launch: the waiting list is open, with founder pricing for the first cohort.

Conclusion

Sovereign company AI is no longer a theory. It is a marten from Burgundy that answers, cites its sources, and never talks nonsense, on an appliance you can unplug. BrainDup is the start of a long series. The next version adds ontology management, followed by a build assistant and a sovereign cloud offering with Infomaniak.

Want to see Marvia at work, or build your own sovereign AI? Explore BrainDup and our training, or book a slot to discuss your project. Join the Chaman d'IA waiting list to be in the first cohort.