New, 4 August 2026
BrainDup v2.2 is out, and Marvia proves it
Marvia, the AI of the Morvan, is the demonstration bench for BrainDup in a fully usable version. 1,423 documents and around twenty websites about the Morvan regional nature park were ingested. Answers stay within the corpus and carry their sources, on a local appliance, with no internet connection.
1 423
documents ingested
29 515
text fragments
30 887
entities in the graph
Measured quality
BrainDup ships with a set of 120 reference questions whose correct answers and correct sources are known, including questions outside the corpus: the ones the system must refuse to answer. Every run is timestamped and kept, append-only.
At the end of July 2026 that check revealed that a maintenance operation had broken the link between documents and their catalogue: the answers were still right, their citations were not. Fixed in production the same day, correct citations up from 17 % to 69 %.
The target remains 100 %. The gap is documented: in most failing cases the information is indeed in the corpus, but the search does not surface the right article. That is the first job on the list.
The vision
Sovereign AI is technically possible, but there was no framework to build it with. BrainDup combines semantic vector search with a structured knowledge graph, reaching 81.67 % accuracy against 57.50 % for vector-only methods.
Proof in the field
MethodMIND validates vector RAG. PST2RAG adds the dual graph schema. IA-ACTES proves production maturity: a legal deed that took seven hours produced in one, with a 96.1 % extraction rate.
The framework today
With Claude Code and the BMAD method, you no longer start from a blank page. A complete development environment, working and extensible, shaped for each new line of business. No stress.
Neither a blank page nor a ready-made app
A complete development environment, working from day one.
A working technical stack
Django, FastAPI, Milvus, Neo4j, Ollama, n8n, Nextcloud. Installed, configured, ready to run.
NVIDIA Grace Ready
Installation tooling for NVIDIA's latest local AI desktops and servers.
Built-in methodology (BMAD and Claude Code)
Nine specialised AI agents guiding every phase of development.
Pre-configured development rules
Code conventions, security, EU AI Act and GDPR compliance, CI/CD pipelines.
Community backing
Living documentation, support, lessons learned from the reference projects.
Proven reference projects
MethodMIND, PST2RAG and IA-ACTES as models: architectures validated in production.
Reference projects
MethodMIND
A semantic search engine over a scientific corpus. 2,000 PubMed abstracts indexed in two hours with vector RAG, SciBERT and T5 reranking.
Résultats
Instant semantic search, with verifiable references and direct links to PubMed.
Enseignement pour BrainDup
Validation of the vector RAG pipeline with Milvus in a demanding setting, where invention is simply unacceptable.
PST2RAG
PST email archives turned into a searchable knowledge base. A pipeline from PST to MBOX to Markdown to embeddings and a Neo4j graph.
Résultats
Semantic search and relationship analysis for compliance, legal e-discovery and digital forensics.
Enseignement pour BrainDup
This is where the Milvus and Neo4j dual schema was born, the architectural pattern that became the core of BrainDup.
IA-ACTES
Automated production of legal documents for pharmacy transfers. Four months of development, 169 merge fields, 12 Docker services, 429 tests.
Résultats
Seven hours per deed brought down to one (a factor of 7). 100 % accuracy on leases, 99.5 % on payslips, validated across 22 pharmacies.
Enseignement pour BrainDup
Proof of industrial maturity: real production, EU AI Act and GDPR compliance, a productivity factor of 2.9 with Claude Code and BMAD.
Take the step
Thirty minutes to assess what sovereign AI can do in your organisation. No commitment.
Free sovereignty review