Transportation and logistics

Driver's "scribbles" are not a sentence

How did artificial intelligence learn to read what ordinary OCR can't see?

TL;DR

  • Traditional OCR works in a template fashion gets lost with crumpled documents or displaced stamps.
  • Modern AI combines handwriting recognition (HTR) with contextual understanding, "guessing" content just like the human brain.
  • Dokuparser automates processes from the photo to the data in the TMS, asking a human to verify only in unclear cases.

Let's talk frankly about what happens on a Friday afternoon in the operations department of a typical transportation company. The shippers are trying to finalize the week, the phones are ringing, and hundreds of photos from drivers are flowing into the email inbox.

And this is where the drama begins. Photos are taken in a hurry, often with a "calculator" in low light in the cabin. CMR documents are crumpled, stained with grease or coffee. But worst of all is what is written on them. Or rather - how they were written. The handwriting of the driver, who filled out the document on his knee, leaning on the steering wheel, is often a puzzle that Enigma cryptologists would not be ashamed of.

If you've ever tried to run such a document through a traditional OCR (Optical Character Recognition) program, you know how it ends. The system spits out errors, bushes and random characters. Instead of automation, you have frustration and manual correction of everything from scratch.

Why is this happening and why does 2026 bring a breakthrough on this issue?

Close-up of a dirty, crumpled delivery document on a vehicle dashboard being scanned by a blue digital laser, illustrating AI handwriting recognition of messy driver scribbles


Why did the old OCR fall in the clash with logistics?

To understand why Dokuparser works differently, we need to explain why the old technology failed.

Traditional OCR (as we have known it since the 1990s) works a bit like a cookie cutter. It expects the letter "A" to always look the same and to be in a perfectly even row. It works great on legal contracts printed in Word. But transportation is not a pharmacy. Here, chaos rules:

  • No template: The recipient's stamp lands where it just happened to be - sometimes in box 24, sometimes in 16, and sometimes in the margin upside down. The old OCR, which searches for data "by coordinates" (Zonal OCR), goes dumb at this point.
  • The human factor: The handwriting is unique. The letters blend together, the "7" looks like a "1" and the signature overlaps the weight of the goods. For a classical algorithm, this is noise, not information.

For years, the IT industry said logistics:"You have to force drivers to write more clearly," he said. Every transportation manager knows that this is utopia. That's why technology had to change.

The HTR and Context Revolution: AI that thinks like a shipper

Modern tools such as Dokuparser, they do not "look" at the document as a collection of pixels. They "read" it - using mechanisms similar to the human brain. This is based on two pillars:

1. HTR (Handwritten Text Recognition)

This is handwriting recognition technology. The AI models have been "fed" millions of writing samples - from doctors' scribbles to school calligraphy. Thanks to this, the system is able to recognize that this strange swirl is the letter "g" and that line over there is the number "1." This isn't guesswork, it's statistics based on neural networks.

2. Semantic Understanding (Context)

This is a real game changer. Imagine seeing a blurry word: "W__szawa." Your brain automatically adds the missing letters because you know it's the capital of Poland. The same way modern AI works in Dokuparser.

  • If the system sees the string "24 00" with the word "kg" or "gross" next to it, it knows it is a weight, even if one zero is blurred.
  • If it sees "12/05/202_" in the date field, it can guess the year based on other documents or the current date.

The system does not need rigid frames. It seeks meaning, not just shapes.

From a photo on WhatsApp to a record on TMS - without the headache

What does this look like in operational practice? The driver finishes the route. He takes a picture of the CMR. He doesn't have to scan it in the office. He sends it as it stands. Dokuparser takes over this file and in a split second does a job that would take a human being a minute of squinting:

  1. Rotates the image and improves contrast (removes shadows from the cabin).
  2. Locates key fields (regardless of where the stamp was stamped).
  3. Reads handwriting and printed.
  4. Transmits ready data (JSON/XML) to your TMS or ERP system.

Importantly - the system is fair. If the confidence of the reading (Confidence Score) is low (for example, an oil stain flooded half the amount), the system does not guess. It highlights the field in red and says:"Hey, man, check it out.". Then your employee corrects only this one passage, and does not rewrite the entire document.

No more "doctoring" on the nature of writing

Implementing this technology is not just about saving time. It's a change in the quality of your team's work. Instead of getting frustrated and calling the driver to ask"What did you write here, Marek!", freight forwarders do what's important - finding cargo and providing customer service.

Technology has finally stopped requiring us to be robots writing in printed letters in boxes. In the end, it's the machine that has learned to understand us - with all our clutter, haste and imperfection.

Do you have a document that every other system has relied on? That's great. We like a challenge. Take that crumpled CMR that's sitting on your desk, take a picture of it with your phone and upload it to the Dokuparser. See for yourself how artificial intelligence handles "scribbles" that until now only their author could read.