Cloud and edge integration for MedTech, designed from the only question that matters when connectivity fails: what can the device still do on its own.
Moving processing to the cloud is usually justified on development speed, maintainability and the ability to improve logic without a firmware release. All of that is true, and each move makes the device dependent on your platform being reachable. For a device in a hospital basement during a network incident, the explanation that the cloud was unavailable is not an acceptable account of why an alert did not arrive. The split between device and cloud therefore belongs to the clinical risk conversation rather than only the engineering one.
The first architecture decision is not which cloud service to use. It is which functions must remain available when the cloud is not.
Document processing for manufacturing where every extracted value is verified against something: the specification, the order, the receipt or the current revision.
A certificate of analysis is not useful because it has been read. It is useful because somebody confirmed that every measured value falls inside the specification for the material you ordered. An invoice is not useful until it matches the order and what was received.
The third column is where the value is in every row. A project that delivers the second column has automated the typing and left the judgement where it was.
Certificates filed rather than checked
Which is a control that exists on paper only.
Data typed from a document into the ERP
Twice, by two people, in two departments.
Documents in four repositories and an inbox
So finding one later is a search.
Every supplier has a different format
And a new supplier breaks whatever was built.
Approvals waiting in somebody’s email
With no visibility of where anything is.
Extraction errors nobody catches
Because there is nothing to check against.
A scan that cannot be read
And no route back to the supplier who sent it.
No link from the value to the document
So the record cannot be substantiated or found later.
Take ten certificates of analysis from your files and check every value against the specification. By hand, properly. In most plants this exercise finds at least one value that should have been questioned, and occasionally one that should have stopped the material. That result is the business case for this page, it takes a morning, and it is considerably more persuasive than a processing time saving.
Document projects usually begin with extraction because it is the visible part. We begin with what each document has to be checked against, because that determines which fields matter, how accurate they need to be, and where the value actually is.
It also determines feasibility: a check against a specification requires the specification to exist as data, and in several plants that is the real piece of work.
Step 1
Assess the document workflow as it runs today, end to end
Step 2
Classify the document types and the volumes behind each
Step 3
Define the check for each type: against what, field by field
Step 4
Confirm the reference data exists as data, not as documents
Step 5
Define the fields to extract, and the consequence of each
Step 6
Design the extraction, with the format tail accounted for
Step 7
Configure and train against your own documents
Step 8
Build the validation and business rules with the owner
Step 9
Design the human review interface for speed, not completeness
Step 10
Integrate with ERP, PLM, QMS and the workflow
Step 11
Pilot on real volume including the awkward suppliers
Step 12
Deploy, then tune the thresholds against real error experience
Ask where the specification lives, and whether it is data or a document. If it is a document, the first phase of this work is making it data, and that is worth knowing before a budget is set. It is also the phase with value beyond this project, because every downstream check and every piece of quality analytics needs the same thing.
Grouped by what each part does. The validation group is what distinguishes this from a faster typing pool, and it is specified per document type with the owner of that document.
Classification and extraction
Intelligent document classification, including the documents that arrive untitled in an inbox
Optical character recognition and structured data extraction, configured against your own document population
Table and line-item extraction, which is where most invoice and order effort actually sits
Handwriting and poor-scan handling to the extent it is reliable, with the limits stated rather than implied
Engineering drawing metadata extraction: part number, revision, title block, which is the part that works
Validation and checking
Certificate of analysis validation against the specification, property by property, with units resolved
Material test certificate validation against the specification and the lot delivered
Three-way matching for invoices: order, receipt and invoice, with tolerance rules you set
Purchase order and delivery document verification against your own order data
Inspection report checking against acceptance criteria
Business rule engine for the checks that are specific to your products, customers and suppliers
Workflow and integration
Automated routing and approval workflows, with visibility of where every document is
Human review interfaces designed for speed, per section 3
ERP, PLM, QMS and document management integration, with the value linked to its source document
Supplier onboarding document verification, including completeness checking of a required document set
Exception queues with an owner and an age, including the documents that cannot be read
Records and analytics
Document repository integration with the original retained, linked to every extracted value, and findable by the people who need it later, which is the traceability and accessibility requirement together
Audit trail: what was extracted, what it was checked against, what the result was, who reviewed it
Supplier document quality analytics, which is the feedback loop in section 5
Straight-through rate and tail analysis by document type and supplier
These are ordered deliberately. The first three create value through checking rather than through speed, and they are where the return is largest even though the volume is often lower.
Certificate of analysis processing
Extracting measured values and verifying each one against the specification for the material ordered.
Material test certificate validation
Verifying test results against the specification and confirming the certificate belongs to the lot delivered.
Supplier invoice matching
Three-way matching against the order and the receipt, with your tolerance rules applied.
Purchase order data extraction
Reading incoming customer orders and verifying them against your agreed terms and prices.
Inspection report digitisation
Converting incoming or legacy inspection reports into data, checked against acceptance criteria.
Engineering drawing metadata extraction
Part number, revision, title block and related metadata, verified against the current revision in PLM.
Shipping and logistics document processing
Delivery notes, packing lists and customs documentation, checked against the shipment and the order.
Supplier onboarding document verification
Confirming a required document set is complete, current and valid before a supplier is approved.
Rank your suppliers by how much document handling they cost you. Once documents are processed with measurement, this list exists for the first time and it is usually concentrated: a small number of suppliers generating most of the manual effort. That is a procurement conversation with evidence, and it is worth more than another increment of extraction accuracy.
We work within your existing systems and document repositories. The architecture is a pipeline, and the design decisions that matter are what each stage does when it is unsure.
Ingestion
Email, portal, scanner and system feeds, with the original retained as received.
Classification
What kind of document this is, with an unknown route rather than a guess.
Extraction
Fields and tables, with a confidence per field.
Language understanding
For the documents where layout rules are not enough.
Validation
Business rules and the checks from section 2, against reference data.
Human review
Fast, keyboard-driven, with the reason for the flag shown.
Integration
ERP, PLM, QMS and DMS, with the value linked to its source.
Records and audit
Original, extraction, check result, reviewer, retained together.
Layout-based extraction where the format is stable and rules are reliable, because a rule that works is cheaper to run and easier to explain than a model
Model-based understanding where formats vary, which is most supplier documents, with confidence used as described in section 2
Language models where a document has to be interpreted rather than read, with the same boundary discipline as our copilot page: grounded in the document, never filling gaps from general knowledge
Reference data read from the system that owns it rather than copied into the document platform, so a specification change takes effect immediately
And no model or document leaving your boundary without your IP, legal and security functions’ determination, which is the same position as the copilot page
Manufacturing documents substantiate decisions: that material conformed, that an invoice was correct, that an inspection passed. That makes them records, and the controls follow from that rather than from the processing.
Access and authorisation
Who may see which documents, including commercial terms.
Sensitive information
Commercial, personal and proprietary content handled per your policy.
Extraction accuracy
Thresholds per field, with checks against reference data.
Version management
Which revision a document is, and which is current.
Retention and traceability
Original retained, linked, and kept per your obligations.
Human review
Required where a value feeds a quality or financial decision.
Secure processing
Where documents are processed and what may leave your boundary.
Change control
Rule and threshold changes through your process, with a record.
Extraction accuracy is the number vendors report and it is not the number that matters. A system can extract every field correctly and still send most documents to a person, because the checks fail or the confidence sits below the threshold.
STRAIGHT-THROUGH RATE by document type and by supplier: the proportion that required no human touch. This is the business case
Where the rest stopped, by reason: low confidence, failed check, unknown format, unreadable. Four different problems with four different fixes
Review time per document, since the benefit is straight-through rate multiplied by what a review costs
Turnaround time end to end, including the waiting, because that is what the business experiences
Check failure rate, separated into extraction errors and genuine non-conformances. The second group is the quality value and should be reported to quality rather than treated as noise
Cost per document against the baseline, which requires the baseline to have been measured first
Non-conformances found by the check that would previously have passed, which is the number that justifies the whole approach
Supplier document quality, per section 5, as a procurement input
Exception queue age and anything unresolved, per section 6
Correction rate after acceptance, which indicates a threshold set too permissively
Extraction accuracy per field, which is a diagnostic rather than a headline
Measure the straight-through rate on the first day and the reason for every stop. Those two numbers direct every improvement afterwards: more training data, a better rule, a threshold change, or a conversation with a supplier. Without the reasons, improvement work goes wherever the last complaint came from, and the obvious wins sit untouched for months.
Outcomes depend on your document volumes, your supplier base and whether your reference data exists as data. These are the changes the work is aimed at.
Certificates are verified against the specification property by property, which in many plants is a control starting to operate for the first time
Invoices are matched three ways with your tolerance rules, consistently
Non-conformances are found at receipt rather than at use
Manual data entry falls, and the time saved is measured against a baseline rather than estimated
Every value in the system can be traced to the document and the check behind it
Documents that cannot be read reach an owner rather than a folder
Supplier document quality becomes a measurable performance attribute
Approval status is visible rather than sitting in somebody’s inbox
And the straight-through rate is known, with the reason for every exception
Bring us one document type and its check, and we will assess the formats, confirm whether your reference data exists as data, define the check field by field, and tell you the straight-through rate you should expect before anybody configures anything.
CaliberFocus builds manufacturing document processing where every extracted value is verified against a source of truth: certificates against the specification, invoices against the order and receipt, drawings against the current revision. We define the check before the extraction, set thresholds by the cost of a wrong value, and design the review interface for speed.
- Certificates checked against the specification, property by property
- Three-way invoice matching with your tolerance rules
- Every value traceable to its source document and the check behind it
- Straight-through rate measured per supplier, with the reason for every stop
Security & Compliance
