IA Act Applicable depuis le 2 août 2026 : les preuves que BrainDup produit. Voir la conformité

Reference projects

MethodMIND, PST2RAG, IA-ACTES: proof in the field

A connector standing for the integration between BrainDup's building blocks

BrainDup is not a theoretical concept. It is a framework that matured through real projects in different domains: scientific research, email archiving, legal document production. Each project put the architecture under new constraints, strengthened the technical base and validated design choices.

What sets BrainDup apart from yet another AI framework is exactly that trajectory: our past projects were not built to illustrate the framework. The framework emerged from those projects, by accumulating proven patterns and lessons from the field.

MethodMIND

Managing scientific documents

Context

Delivered as part of the Le Wagon bootcamp in Barcelona, batch #1829, October to December 2024. A team of five developers.

The problem

Neuroscience researchers spend weeks reading papers to identify validated experimental methods. General-purpose language models invent: they produce references that do not exist. For a researcher, a fabricated citation is a risk to the integrity of their work.

The solution

  • Corpus: 2,000 PubMed abstracts indexed in two hours
  • Embeddings: SciBERT (768 dimensions), a model built for scientific language
  • Vector store: Milvus, for similarity search at scale
  • Reranking: T5 MonoT5-MSMarco, to sharpen relevance
  • Security: an anti-injection filter based on the corpus keywords
  • Interface: Streamlit

Technical pipeline

User Input → Firewall → Query Enhancement → SciBERT Embedding → Milvus Similarity Search → PubMed Retrieval → T5 Reranking → LLM Generation

Results

Instant semantic search with verifiable references. Every answer includes the source abstracts and direct links to the original articles on PubMed.

What BrainDup learned from it

MethodMIND validated the vector RAG pipeline with Milvus in a demanding setting: a large corpus, expert users, and zero tolerance for invention. The first project, and the core the BrainDup architecture grew from.

PST2RAG

An AI archivist for email

The problem

Years of organisational knowledge lie buried in email archives. PST files (Microsoft Outlook) hold the history of exchanges, decisions and relationships, but that proprietary format is entirely unreadable to modern AI.

The solution

Extraction and transformation:

  • PST to MBOX (via libpst/readpst) to Markdown
  • Entity extraction (people, organisations, attachments)
  • Intelligent chunking, fitted to the structure of email

Dual indexing:

  • Milvus (semantic search, 384-dimension embeddings): finds content by similarity of meaning
  • Neo4j (relationship graph): models the structural links between Email, Chunk, Person, Entity, Attachment and Organization

Business uses

  • Compliance and e-discovery: quickly find the relevant exchanges in a regulatory or litigation context
  • Digital forensics: reconstruct communication chains and spot patterns
  • Knowledge management: put an organisation's decision history to work
  • Business insight: map networks of influence and flows of information

Results

Semantic search works fully: “budget concerns” does find “financial worries”. Relationship analysis works: “who talks to whom, about what?”. Hybrid queries combine vector search with graph traversal.

What BrainDup learned from it

PST2RAG introduced the Milvus and Neo4j dual schema, the architectural pattern that became the core of BrainDup. Vectors for meaning, the graph for structure. Both together for reliability.

IA-ACTES

Producing legal documents

The problem

Drafting a pharmacy transfer deed takes seven hours per file. 169 fields to fill by hand. 103 documents to process per file. 35 conditional blocks to assess depending on the case. It is expert work, repetitive yet exacting, where every error carries real legal consequences.

The solution

Technical stack:

  • Backend: Django, FastAPI and n8n (orchestrating 84 automated tasks)
  • AI: Milvus, Neo4j and Ollama (local inference)
  • Documents: Nextcloud, Collabora and Docling on GPU
  • Infrastructure: NVIDIA Blackwell; no data leaves the server

Dual-schema RAG:

  • A “Case” schema: the client's case documents
  • A “Legal corpus” schema: the statutory reference texts
  • Three query modes: within the case, within the corpus, or across both

Compliant with the AI Act and the GDPR: human confirmation is mandatory at every critical step.

Results

IndicatorValue
Time per deed7 hours to 1 hour (a factor of 7)
Document extraction rate96.1 % across 56 test documents
Automated merge fields169 (43 % through public APIs, 57 % through AI)
Documents handled per file103
Automated tasks84
Generation modelGLM-4.7-Flash (30B-A3B)
Embedding modelE5-Mistral-7B (4096 dimensions)

Scale of the project: 124 pages of specification, 11 EPICs, 35 functional and 10 non-functional requirements. Built entirely with Claude Code and the BMAD method.

What BrainDup learned from it

IA-ACTES is the proof of industrial maturity. No longer a prototype, but a production system handling real legal documents, under regulatory constraints and with a measurable productivity gain. IA-ACTES validates the whole: the BrainDup technical architecture, the BMAD method, and Claude Code as the primary development tool.

Marvia

The demonstration bench for BrainDup v2.2, revealed on 4 August 2026

The problem

How do you prove a company AI keeps its promises without exposing a client's sensitive data? By having it learn an entire territory. Marvia is a marten, the emblem of the Morvan regional nature park, and BrainDup's public test bench.

The solution

A corpus of 1,423 documents and around twenty websites about the Morvan, ingested and then queryable in plain language, on a local appliance with no internet connection.

  • 29,515 text fragments indexed for hybrid search, semantic and lexical
  • 30,887 entities linked in the knowledge graph, explorable by clicking
  • A sourced answer in 20 seconds with graph traversal, 2.5 seconds in concise mode

See the Marvia demonstration →

In summary: progress that accumulates

MethodMIND laid the foundations. The vector RAG pipeline with Milvus, semantic search over a large corpus, the filter against invention.

PST2RAG added the relationship dimension. The Milvus and Neo4j dual schema turns content search into genuine documentary insight.

IA-ACTES proved production maturity. The complete architecture, deployed on a demanding business case, with measurable results and regulatory compliance.

That progression, vector RAG, then a dual schema with a graph, then production maturity with the BMAD method, then public proof with Marvia, was not planned. It emerged from the field. And that is exactly what makes BrainDup solid.

Take the step

Thirty minutes to assess what sovereign AI can do in your organisation. No commitment.

Free sovereignty review