← All case studies
CASE STUDY · SHIPBUILDING · TRACIUM

Supplier manuals, turned into ERP-ready maintenance plans.

Shipyards receive a manual from every equipment supplier. For Tracium, our sister company, we built an AI pipeline that turns each manual into an ERP-ready maintenance plan, citing the source PDF and page for every value, plus a RAG chatbot for questions. We delivered it in two phases: an MVP on AWS, then production on an NVIDIA GPU server in the customer’s environment.

70% less manual documentation10,000+ technical documentsOn-prem LLM in production

UPDATED OCTOBER 2026

The challenge

Every piece of equipment on a ship arrives with its supplier’s manual: long, inconsistent documents with tables, diagrams, part references and revision history. Maintenance plans had to be found and typed into the ERP by hand, which slowed delivery and made it hard to prove which page a value came from.

Two phases, from cloud to on-premises

PHASE 1 · MVP ON AWS

Less sensitive documents, in the cloud

The MVP runs on AWS with less sensitive documents, using Claude on Amazon Bedrock. Its architecture is shown below.

PHASE 2 · PRODUCTION ON-PREMISES

Sensitive documents, on the customer’s own GPU server

Production runs on an NVIDIA RTX PRO 6000 Blackwell GPU server in the customer’s environment. We deployed a self-hosted open-source LLM on it, so sensitive documents are processed on-site.

How we run AI on-premises →

What we delivered

Document ingestion

PDF and scanned-document processing, OCR where required, metadata capture, versioning, and durable source storage.

Structure-aware extraction

Content is segmented using document hierarchy and layout so tables, sections, and cross-references retain useful context.

Retrieval and grounded answers

A RAG chatbot lets engineers ask questions about any manual. Embeddings and vector retrieval find the relevant passages, and Claude answers with a citation to the PDF name and page.

Structured outputs

Each supplier’s maintenance plan, with its tasks, intervals and spare parts, is extracted into a fixed schema and delivered as a structured document the ERP can import, with every value citing the PDF name and page it came from.

Evaluation and review

Every extracted field gets a confidence score. Low-confidence fields go to a person for review before anything reaches the ERP, and test sets measure extraction accuracy.