Customer story · Aviation
A regional airline
Invoice processing and cost analytics for a regional airline
Finance uploads invoices in bulk. Nexus extracts, classifies and validates them, and the team gets a cost portal and an assistant that answers questions from the extracted data.
Executive summary
The situation
Scans, stamps, foreign-language invoices and inconsistent supplier names made the document set difficult to use as a coherent source of cost information. Currency and payer identity also mattered to a meaningful total.
What was built
Bulk upload feeds native extraction with vision/OCR fallback, classification, arithmetic checks, duplicate removal and vendor mapping. Structured data supports a finance portal, reports and a grounded cost assistant.
Why it matters
Makes invoice information easier to browse and question without treating every extracted number as comparable. Rules preserve currency boundaries and distinguish the airline’s payables from other amounts.
- Primary users
- Finance staff
- Core systems/data
- Supplier invoices, scans and cost data
- Human control point
- Finance inspects documents and vendor mappings
- Nexus role
- Extraction, classification and validation
The challenge
Invoices that were never meant to be read by a machine.
The invoice set was not a clean dataset. Scanned PDFs, stamps and foreign-language documents complicated reading, while supplier names varied between invoices. Even correctly extracted text could leave the same vendor represented by several identities.
A cost picture needs more than text extraction. Categories, arithmetic checks, duplicates, currency and payer identity all affect how the figures can be used. Those finance distinctions shaped the pipeline as much as the document-reading task.
What we built
A document pipeline with a cost portal on top.
Finance uploads invoice batches through a web interface. Nexus first attempts native text extraction, then uses vision and OCR where a scan needs that fallback. The information moves through cost classification, arithmetic checks and duplicate removal before being stored as structured invoice data. Reading the document is the start of the process, not the end.
Vendor canonicalisation addresses a different problem: spelling variants should not become separate supplier identities. The portal brings cost KPIs, document browsing and vendor mapping together with a structured report. A grounded assistant answers questions from the extracted data, subject to rules that prevent cross-currency summation and only count payables where the airline is the payer.
How the workflow runs
Step by step.
01
Bulk upload
Finance uploads a batch of invoices through the web interface. Native text extraction first, with vision and OCR fallback for scans and stamped documents.
02
Classify
Each invoice is assigned to a cost category. Arithmetic checks confirm line items and totals; duplicates are removed.
03
Canonicalise vendors
Supplier name variants are mapped to one vendor identity. Structured invoice data feeds the cost portal, reports and assistant.
What the business team sees
A cost portal for finance.
The portal lets finance move between a cost question and its underlying documents. Supplier and category views sit alongside document browsing and vendor mapping. The assistant provides a conversational route into that data, but asking for a total does not make amounts in different currencies suitable for addition.
- Cost KPIs by category
- Document browser with extracted fields
- Vendor mapping
- Structured invoice report
- Grounded cost assistant
How Nexus works underneath
Reading documents is only the first part.
Nexus connects native extraction or vision/OCR to classification, deterministic arithmetic checks and duplicate removal. Supplier identity is a deliberate part of the model, not an accidental consequence of invoice spelling. AI helps read and query the material; rules determine how the figures may be combined, including in the assistant’s answers.
Architecture
How the layers fit together.
Business team
Upload interface and cost portal
Bulk upload · KPIs · document browser · vendor mapping · assistant
Nexus orchestration
- 01Native extraction
- 02Vision/OCR fallback
- 03Cost classification
- 04Arithmetic validation
- 05De-duplication
- 06Vendor canonicalisation
Connected systems and data
- Invoice PDFs and scans
- Cost database
- Reports
Human control and audit
Answers grounded in the extracted data.
The assistant answers from the structured invoice data, not from general knowledge. Domain rules constrain how figures are combined, so a question about total spend does not silently mix currencies or include documents where the airline is not the payer.
What this changes
Documents become a working dataset. Gives finance structured records and a document browser alongside the invoice material.
Supplier identity is explicit. Canonicalisation is designed to keep name variants from fragmenting vendor reporting.
Questions retain their context. The assistant can help explore cost without ignoring currency or payer boundaries.
Your process
Discuss a similar workflow
Tell us about the documents, systems and decisions involved. We will discuss where Nexus orchestration fits and where your team stays in control.