Building the AI Foundation That Raised €2M
How a 4-month proof-of-concept for maritime AI became the foundation for a startup's seed round. RAG pipeline, knowledge extraction, and the architecture decisions that made it investable.
In early 2024, I led a small team building a proof-of-concept: an AI-powered knowledge base for the maritime industry. Four months later, the PoC became a funded startup, raising over €2M in seed.
The Problem
Maritime companies sit on decades of unstructured knowledge: safety procedures, vessel specifications, regulatory documents, operational manuals. When someone needs an answer, they either ask the one person who remembers, or they dig through folders of PDFs.
The client wanted to know: can AI make this searchable and useful?
What We Built
A RAG (Retrieval-Augmented Generation) pipeline purpose-built for maritime documents:
- Document ingestion: PDF parsing, chunking strategies tuned for technical manuals (headings-aware, table-preserving)
- Vector search: embedding pipeline with hybrid retrieval (semantic + keyword)
- Domain-specific prompting: maritime terminology, regulatory references, safety-critical answer formatting
- Source attribution: every answer traces back to the specific document and section
The stack was TypeScript end-to-end: Node.js backend, React frontend, PostgreSQL with pgvector for embeddings.
The Hard Parts
Chunking maritime documents is not like chunking blog posts. A safety procedure might span three pages with numbered steps, cross-references, and embedded tables. Naive chunking by token count breaks the semantic units. We built a heading-aware chunker that preserves document structure.
Accuracy is non-negotiable in maritime. A wrong answer about a safety procedure could have real consequences. We built a confidence scoring layer: the system knows when it’s guessing and says so. Answers below threshold show a warning and link directly to the source document.
The team was small. Four engineers, four months, shipping weekly. I led architecture decisions and owned the RAG pipeline. The PoC needed to be good enough to demonstrate value, but architected well enough that a real engineering team could scale it.
The Outcome
The PoC demonstrated clear value: domain experts validated that answers were accurate, retrieval was fast, and the interface made decades of institutional knowledge accessible in seconds.
The client used the PoC to raise a €2M+ seed round. The startup now has its own engineering team, building on the architecture we designed.
What I Took Away
Building a PoC that becomes a funded company is different from building production software. The architecture has to be investable: clean enough that incoming engineers can understand it, extensible enough that investors see a product roadmap, and solid enough that the demo doesn’t break mid-pitch.
The maritime domain also reinforced something I’d seen before: professional industries have massive amounts of trapped knowledge. The tooling to unlock it is mostly the same (RAG, extraction, structured search), but the domain-specific tuning is where the value lives. Generic AI doesn’t cut it. You need to understand the documents.