OCR Solutions
Paper turned into data, with the numbers checked harder than the words.
What is OCR solutions?
OCR converts text in images and scans into structured data your systems can use, extracting named fields from forms, receipts and documents rather than simply producing a wall of text.
Reading text from a clean PDF is close to solved. Reading a phone photograph of a creased delivery note that spent the day in a vehicle is a different problem, and that is the one most businesses actually have.
The distinction that matters most is between printed and handwritten. Printed text is reliable enough to automate; handwriting is substantially harder and anybody quoting the same confidence for both is not being straight with you.
There is also an asymmetry people miss. A misread letter in a customer name is a nuisance. A misread digit in an amount is a financial error, so numbers deserve stricter checking than words.
Why numbers are treated more strictly than words
Duara builds OCR from Nairobi for businesses still moving information off paper, including delivery notes, receipts, forms and supplier invoices.
We set stricter confidence thresholds on numeric fields than on text, because the cost of a wrong digit and a wrong letter are not remotely the same. Systems treating every field identically look simpler and produce expensive errors in exactly the fields that matter.
Who this is for
- Data is typed in from paper forms
- Delivery notes arrive as photographs
- Receipts are captured manually for expenses
- Archives exist on paper and cannot be searched
Industries we serve
- Logistics and distribution
- Financial services and lending
- Insurance claims
- Healthcare records
- Government suppliers
- Retail and expenses
Problems this solves
These are transcription tasks, which is where OCR earns its place when the source quality allows.
- Typing from paper. Staff retype information that already exists in a document, which is slow and introduces errors at the point of entry.
- Photographs instead of scans. Documents arrive as phone images taken in poor light, which is the real input condition rather than the ideal one.
- Paper archives that cannot be searched. Historical records exist and finding anything in them requires somebody physically looking.
- Errors found late. A mistyped figure is discovered during reconciliation, long after the cheapest moment to catch it.
What we deliver
Field level extraction from your actual documents, with confidence thresholds set higher for numbers than text, uncertain fields routed for checking, and accuracy measured per field rather than overall.
Field level extraction from your document types
Stricter confidence thresholds on numeric fields
Uncertain fields routed for human checking
Accuracy measured per field, not just overall
Integration into your existing systems
Searchable text from paper archives
Features
- Numbers checked harder than words. A wrong digit in an amount is a financial error where a wrong letter in a name is a nuisance, and thresholds reflect that.
- Accuracy reported per field. An overall figure hides that the total is read well and the date badly, which is exactly what you need to know.
- Uncertain fields go to a person. Low confidence extraction is queued for checking rather than written into your records as fact.
- Honest about handwriting. Printed text automates reliably and handwriting does not. We quote them differently because they are different problems.
- Tested on your real documents. Assessment uses the creased photographs you actually receive, not a clean sample.
Benefits
- Typing disappears. Information already captured on paper stops being entered by hand a second time.
- Errors are caught at entry. A flagged uncertain field costs seconds; the same error found at reconciliation costs far more.
- Archives become searchable. Paper history turns into something findable rather than something somebody has to look through.
- Processing keeps pace. Document volume stops being limited by how fast people can type.
Technologies we use
- Field extraction. Pulling named values such as totals, dates and reference numbers rather than producing undifferentiated text.
- Confidence scoring. Assigning certainty per field so thresholds can differ between numbers and text.
- Image preprocessing. Correcting skew, contrast and rotation before reading, which materially affects accuracy.
- Validation rules. Checking extracted values against expected formats and against your own records.
- Verification queues. Routing uncertain fields to a person with the image alongside for quick confirmation.
Source quality sets the ceiling and no model recovers what the image never captured. Improving how documents are captured is often cheaper than improving how they are read.
How the process works, step by step
An OCR project usually takes four to eight weeks, and it begins by testing on documents in the condition you actually receive them.
-
Step 1
Free document assessment
We test on your real documents, including the poor quality ones, since those determine the outcome.
-
Step 2
Capture improved first
Where scanning or photography can be improved, that is done first because it is cheaper than compensating for it.
-
Step 3
Fields and thresholds agreed
Which fields matter and how strict each should be, with numbers set tighter than text.
-
Step 4
Verification workflow
The checking queue is built so uncertain fields have somewhere to go before they reach your records.
-
Step 5
Measured per field
Accuracy is reported field by field, and we say plainly where it is not good enough to automate.
Quality assurance and how we work
The commitments attached to turning documents into records people will rely on.
- Numeric fields get stricter thresholds. The cost of a wrong digit and a wrong letter are not remotely the same, and treating every field identically produces expensive errors.
- Accuracy is reported per field. An overall percentage hides that the total reads well and the date does not, which is the thing you actually need to know.
- We are honest about handwriting. Printed text automates reliably and handwriting does not. Quoting the same confidence for both would not be straight with you.
- Uncertain values never enter records silently. Low confidence fields go to a verification queue rather than being written in as fact.
- We test on your worst documents. Assessment uses the creased photographs you actually receive, because a clean sample proves nothing about your operation.
Frequently asked questions about OCR solutions
How accurate is OCR on our documents?
It depends entirely on their condition, which is why we test on your real ones including the poor quality examples before quoting.
Can it read handwriting?
Less reliably than printed text, and we say so. Anybody quoting the same confidence for both is not being straight with you.
Why treat numbers differently from text?
Because a wrong digit in an amount is a financial error and a wrong letter in a name is a nuisance. The thresholds should not be the same.
What happens to uncertain fields?
They go to a verification queue with the image alongside, rather than being written into your records as fact.
Can it read phone photographs?
Yes, and that is the real input for most clients. Quality varies and preprocessing helps considerably.
Would better scanning help?
Usually more than better software. No model recovers what the image never captured, so we improve capture first where we can.
Does it work on Kenyan documents?
Yes, including delivery notes, receipts and forms. We test on your specific layouts rather than assuming a standard.
Can it read faded thermal receipts?
Partially, and it is one of the harder cases. The assessment tells you honestly whether it is good enough to automate.
How do you report accuracy?
Per field rather than overall, because an average hides that the total reads well and the date does not.
How long does an OCR project take?
Four to eight weeks, starting with an assessment on documents in the condition you receive them.
Can it handle several document layouts?
Yes, though each layout adds work, and we would ask whether all of them genuinely need automating.
Where do the extracted values end up?
Written straight into your ERP, accounting or records system, rather than as files somebody then rekeys.
Can it make old archives searchable?
Yes, and this is often the quickest win, since finding a document is usually the immediate pain rather than data entry.
What about ID documents?
Possible, and it raises data protection questions we would want to scope carefully before building.
Does it validate what it extracts?
Yes, against expected formats and against your own records where they exist, which catches errors confidence scores miss.
Will it replace our data entry staff?
It removes the typing. Most clients move those people to verification and exception handling, which needs judgement.
Can documents be processed without leaving our building?
Yes, on site processing is possible where records cannot leave the premises, which matters in some regulated work.
What if accuracy is not good enough?
We tell you at assessment rather than after the build. Some document conditions do not support automation and it is better to know.
Where are the documents stored?
In your systems. Where processing runs on site, images need not leave the building at all.
What does an OCR project cost?
Document volume and layout count determine it, quoted after the document assessment.
Glossary of terms
Document capture terms in plain language.
- OCR
- Optical character recognition, the reading of text from an image or scan.
- Field extraction
- Pulling specific named values from a document rather than all of its text.
- Confidence score
- How certain the system is about a particular extracted value.
- Verification queue
- Where uncertain fields go for a person to confirm against the image.
- Preprocessing
- Correcting skew, contrast and rotation before reading, which materially affects accuracy.
- Structured data
- Values in named fields your systems can use, as opposed to a block of text.
Related services in AI Development
Talk to us about OCR solutions
Send us thirty documents in the condition you actually receive them, including the bad ones. That assessment answers the question.
Start your project