AI · Computer Vision & Document Intelligence
A picture becomes a record, and the record says how sure it is.
Noble builds systems that read what a person would otherwise have to look at: the invoice photographed on a phone, the form scanned at a counter, the identity document held up to a camera, the component on a line inspected against a drawing, the site checked against what was there yesterday. Every output is a structured record with a confidence on each field, and everything below the organisation’s threshold goes to a person whose correction goes back to the model.
Why documents are the harder problem
A camera on a production line sees the same kind of thing all day under the same light, and a model can be trained to it. A document pipeline sees whatever arrives: a supplier’s invoice in a layout the model has never met, a form filled in by hand in Arabic with the amount written in Western digits, a contract scanned at an angle with a stamp across the clause that matters. Document intelligence is not recognising characters — that part is solved — it is understanding structure: which number is the total, which table row belongs to which line item, which of the three dates is the one the process needs.
In this region that problem has a second dimension. Arabic documents run right to left, their tables do too, and a great many real documents mix the two scripts on one page — an Arabic form with English product codes, a bilingual contract with clauses side by side. Noble treats that as the normal case rather than the edge case, and the evaluation set for any document system here contains the organisation’s actual documents in both scripts, at the quality they actually arrive at.
Many inputs, one pipeline
Whatever arrives — a scan, a phone photograph, a camera frame, a native PDF — converges on the same stages: understand the structure, extract the fields, attach a confidence, and route what is uncertain to a person.
- 01Scanned form or contractArabic, English, mixed
- 02Photographed documentFrom a phone, at an angle
- 03Camera frameLine, site, gate
- 04Native PDF with tablesStructure, not just text
- 05Text and image togetherA claim with its photographs
What the service covers
Extraction from documents that arrive as they are
Invoices, purchase orders, delivery notes, forms, identity documents, permits and contracts read into structured fields — with layout understood rather than assumed, so a supplier’s new invoice template does not break the pipeline. Extraction models are chosen per document class and evaluated on the organisation’s real documents before anything is connected to a process.
- Invoices, orders, forms, identity documents and contracts
- Layout understood, so new templates do not break the flow
- Handwriting in Arabic and English
- Evaluated on the organisation’s real documents first
Document understanding — structure and relationships
Beyond the fields: which line items belong to which order, which clause references which schedule, whether the total agrees with the sum of the lines, whether the signature block is complete. Tables are read as tables, including right-to-left ones, and cross-references inside a document are resolved so that a downstream process gets a record it can act on rather than a bag of strings.
- Tables read as tables, right-to-left included
- Line items, totals and cross-references resolved
- Validation rules applied before the record is released
- Bilingual documents handled as one page, not two
Visual inspection against defined criteria
Classification and detection on images from a fixed camera or a handheld one: the component that does not match the drawing, the label that is missing, the corrosion that was not there last month, the safety equipment that is not being worn. The criteria are defined with the people who inspect today, and the system flags for their judgement rather than replacing it.
- Defect, absence and change detection on fixed and handheld cameras
- Criteria defined with the people who inspect today
- Site and asset comparison over time
- Flags for a person’s judgement, not in place of it
Multimodal — text and image together
The cases where the picture and the words arrive as one thing: a warranty claim with its photographs, an inspection report with the images it describes, a shipment with its documents and the photograph of the pallet. Handled as one record, so that the claim can be checked against the picture and the picture against the claim.
- Claims, reports and shipments as one record
- Text checked against image, and image against text
- Vision-capable language models applied where they fit
Confidence, thresholds and the person
Every extracted field carries a confidence, and the organisation sets the threshold below which a person looks. What that person corrects goes back into the evaluation set and, where appropriate, into the model, so the threshold can be moved with evidence rather than optimism. The review screen is designed for the reviewer — the source shown beside the field, the uncertain part highlighted, one keystroke to accept.
- A confidence on every field; a threshold set by the organisation
- Corrections returned to the evaluation set and the model
- Review screens designed for the reviewer
- Thresholds moved on evidence
Into the process, not into a folder
An extracted record is worth nothing in a spreadsheet. The pipeline posts to the system the process runs in — the ERP, the claims system, the case file — through its released interfaces, with the source image attached and the confidence carried, so the person approving the transaction can see where each number came from.
- Posted to the system of record through released interfaces
- Source image and confidence carried with the record
- Retention and residency for images decided in advance
Where this connects
- SAP Automation, AI & Intelligent EnterpriseWhere the process is inside SAP, document intelligence posts to the standard transaction and is described there from the SAP side.
- Machine Learning & Predictive IntelligenceThe evaluation discipline — held-back data, expected failures, advisory first — is the same, applied to a different kind of model.
- AI Engineering, MLOps & GovernanceImage pipelines, serving cost and the retention rules for what a camera captured are that page’s subject.
The services alongside
- SecurityIdentity documents and personal images are personal data; their handling, residency and retention are set with that service before a pipeline is built.
- Software & Web EngineeringThe capture application — the phone camera flow, the counter scanner, the review screen — is built by this service.
- Managed IT ServicesCameras, scanners and the network they sit on are physical infrastructure, supplied and supported through Managed IT Services and the product catalogue.
Send fifty real documents.
The ones that arrive as they arrive — the crooked scan, the handwritten form, the bilingual contract. Fifty of them, with what the right answer is for each, is the first evaluation set and the honest start of the work.