Before People retype scanned claims field by field.
After A folder read in about 90 seconds for about $0.55, every field with how sure the system is.
Before Quote requests arrive as free-form emails in many languages.
After A filled quote form in about a second, contradictions flagged.
Before Someone checks every machine against its plan by eye.
After One photo, a verdict per slot. Unsure cases say so.
Before Fragments of broken tempered glass are counted by hand.
After The count comes from the photo, scratches ignored.
Before Planning one truck load took over half an hour.
After A load plan in 7–10 minutes, following the stacking rules.
Documents and emails, quality checks from photos, load planning, assistants that talk to your customers: these are only examples, and I am happy to take on a different challenges. I start from what has to get better in your process and deliver a system that does it in 2 steps: a proof of concept in weeks to prove that it works, then a smooth transition to a production ready service running in your AWS account.
Client names are withheld and the pictures are drawn from scratch with fictional data. The problems and the systems are real. Production systems usually start as a proof of concept, so the numbers come from those runs or from tests before go-live; each project page says which.
DocumentsHealth insurance
The system reads the scanned claims that people used to retype by hand. Next to every value it shows the page it was read from and how sure it is, so reviewers start with the doubtful fields.
Chat salesOnline classifieds marketplace
It runs the sales conversation from the reminder about an expiring package to the payment link. It answers from the marketplace rules and the seller's account, and takes the price from the sales system.
Computer visionVending
One photo is enough to check a machine against its plan, without comparing shelves by eye. The system names a product only when it is sure, and otherwise says “not sure”, because a false alarm costs a technician’s trip.
EmailsAir freight software
The system reads emails with a model running in the client's own AWS account in the EU and fills the quote form in about a second. When an email contradicts itself, the form says so instead of quietly picking one number.
Documents and photosIndustrial manufacturing
The compliance card fills itself from photos and the mill certificate, on the Excel template the team already uses, and supplier offers are compared with the request automatically. When the heat number stamped on the steel cannot be matched to the certificate, no card is produced at all.
Scans, photos, email attachments and handwriting. The system pulls out the fields your team retypes today and says how sure it is about each one, so reviewers start with the doubtful ones instead of retyping everything.
See a document projectWhat I know best
Shorter stories, the same approach: a concrete problem, a system that removes it.
Optimisation
The planner uploads a spreadsheet of parcels and the truck dimensions, and the system lays out the load by strict rules. Large, heavy parcels go at the bottom, a larger parcel never sits on a smaller one, and parcels do not overhang each other, so the base sets the outline of the whole stack. Smaller parcels stay in rows, so they do not end up on top of others and a warehouse worker can take them out easily; a 3D view shows the result.
Planning time per load before and with the system, as estimated in the project. Computing cost under $1 per plan.
Computer vision
Quality control breaks a test pane and counts the fragments in a marked square. A segmentation model with classic image processing counts them from a photo and does not mistake scratches for cracks, which a ready-made tool did.
The model was fine-tuned on about 700 photos and checked on separate photos it had not seen before.
Scans, forms, invoices and email requests turned into data your systems use, with a confidence your team can act on. This is my main specialisation.
Quality checks and audits from photos: counting, detecting, comparing a shelf with its plan. Designed so that “not sure” is an honest answer instead of a false alarm.
Planning tasks that eat hours of an expert’s day, such as loading or scheduling, turned into a tool that respects your real constraints.
Assistants that know your products and your customer’s data, answer from your knowledge base and lead to an action: an order, a booking, a callback.
A short call and a sample of the real data, including the ugly cases. You hear what looks feasible and what does not.
A working system on your data in a few weeks, so the decision to go further rests on what it actually did.
The proof of concept becomes a service your systems call, with monitoring and a known cost per case.
Documentation, runbooks and automated tests, so your team can run and change it without me.
Tell me what the problem is and what you expect from the result, and I will see what can be done.
Write to megrzegorz.zyskowski1@Grzegorz Zyskowski
I design and build AI systems end to end: the model, the code around it and the AWS infrastructure that runs it. I work directly with the people who own the problem.
5 AWS certifications, including Solutions Architect Professional and Machine Learning Specialty.