Intelligence
AI in every layer of the system
We connect cyber-physical systems with foundation models, fine-tune them for specific uses, and combine them with deep learning, statistical and fuzzy systems — from the endpoint to the edge, the fog and the cloud.
Transformers
Foundation models & LLMs
Large language and vision models connected to live operational data. Retrieval over your own archive, local deployment, and agents that act inside the system — with a person in the loop.
Adaptation
Fine-tuning for specific uses
Foundation models adapted to a port’s vocabulary, documents and rules — from small task models on the edge to adapters on open-weight LLMs.
Reasoning
Fuzzy systems
Decisions where rules are vague and data is noisy: congestion, priority, risk — expressed the way operators actually think.

TrOCR · VLM
Transformers that read and see
TrOCR for handwriting and print, vision-language models for photos, drawings and screens — every block, line and word located.
Time series
Deep learning on sensor data
Forecasting arrivals, queues and throughput; detecting anomalies in vibration, temperature and movement.
Inference
Statistical models
Estimation, uncertainty and confidence for every number the system reports — so decisions rest on more than a single prediction.
Operations research
Optimisation
Berth, crane, yard and gate planning that turns live data into better moves, not just better charts.
TinyML · edge inference
Endpoint & edge AI
Models compressed to run on trackers and HMI monitors, deciding locally in milliseconds.
Language models, wired into the operation
Operators choose a local or a remote model, ask in plain words, and get answers grounded in live sensor data, the twin and the document archive — deep search, analysis and step-by-step reasoning, with a person deciding.
- Fine-tuned on the port’s vocabulary, documents and procedures
- Retrieval over live data and your own archive
- Local models at the edge when data must not leave the site
- Agents that act through the system, with approval steps
TrOCR · Fine-tuned LLM · LLM × OCR
Handwriting and print recognition with TrOCR, language models fine-tuned on your own documents, and the two joined: OCR that understands what it reads, fills the fields and checks the numbers.
- TrOCR
- Transformer OCR fine-tuned on French and English handwriting and print
- Layout
- Blocks, lines and words located with bounding boxes
- Fine-tuned LLM
- Models adapted to your terminology and document types
- LLM × OCR
- Fields extracted, cross-checked and explained — not just transcribed
- Checks
- Totals, codes and dates validated; low confidence sent to a person
- Deployment
- On-premise — documents never leave your network

Six ways a machine learns to read
TrOCR — handwriting that reads
Transformer OCR fine-tuned on French and English handwriting: forms, invoices, registers, signatures and stamps.
Fine-tuned language models
Domain adaptation on your data and terminology — from small task models to adapters on open-weight LLMs, deployed inside your network.
LLM × OCR
Recognition and reasoning together: fields filled, totals checked, codes validated, and every answer traced back to the page.
Document detection
Bounding boxes for blocks, lines, words, tables, signatures and stamps; deskew and clean-up before reading.
Smart workflows
Documents routed, checked and approved by agents that know when to stop and ask a person.
Living documentation
Procedures, reports and technical files written from the work itself and kept current.
One invoice, read field by field
A handwritten French purchase invoice: every field located, read, checked against the rest of the page — and the one it isn’t sure about sent to a person.
Purchase invoice
Facture d’achat
Invoice no.
F-2026-04192Supplier / Fournisseur
Électronique Exemple3Items
PCB ×40, sensors ×1204Total TTC (€)
4 380,005Date
2 oct. 20266Signature
R. Ka—7SCAN 300 DPI · PAGE 1 / 2
Document report
TrOCR → LLM → rules · fr / en
Specimen QA-DOC-0419 · purchase invoice · illustrative
- [1]RECOGNIZEDDocument typePurchase invoice · Facture d’achat0.97
- [2]RECOGNIZEDInvoice no. (handwritten)F-2026-04190.94
- [3]RECOGNIZEDSupplier (handwritten)Électronique Exemple SARLmatched to vendor record0.88
- [4]INFERREDCost centreR&D · Francefrom line items: PCB, sensors0.79
- [5]VERIFIEDTotal4 380,00 € TTClines + 20 % VAT add up—
- [6]RECOGNIZEDDate2026-10-02 · 2 octobre 2026French date → ISO0.93
- [7]FLAGGEDSignaturePresent · low matchrouted to a person0.42
Diagnosis
Six of seven fields accepted. Line items add up to the total and the supplier matches an existing vendor record. Signature confidence below threshold — sent to a person for review, not guessed.