Paulius Sutkus

Paulius Sutkus, Kaunas. Automation and AI for manufacturing.

I build systems that make everyday work easier.

I build AI and automation systems that connect documents, manufacturing and data into clear workflows: from the first pilot to a working integration.

Selected work

Recent projects

Selected examples from AI, manufacturing automation and infrastructure work. I’d be happy to share other projects and technical details in a conversation.

01 / XMLAdapted to your businessDocuments · accounting · production
PDF / OCRXMLExcelPragma

Document extraction service for businesses

Invoices and production documents → validated data for your software

We extract document fields and line items, highlight total mismatches and prepare data for export. Accuracy is assessed using your sample documents.

  1. Upload a document
  2. Review the data
  3. Export to your software
Document and extracted data

The original invoice beside extracted fields and line items, ready for review and correction.

Export and integration

The working Pragma integration prepares an XML file for import. Data formats and field mappings can be adapted to other accounting or internal software, including Excel (XLS/XLSX) or CSV exports.

Beyond invoices

The solution can be adapted to production documents: orders, job sheets, delivery notes or quality reports. Required fields and validation rules are agreed around your process.

Line grouping and item codes

Several invoice lines are grouped into one Pragma line and assigned a selected item code.

Total checks before export

The system highlights differences between line totals and invoice totals so they can be corrected before XML export.

What the data looks like in Excel

Data from the same demonstration invoice: 9 line items, item codes, quantities, prices and VAT. Totals match the application screenshot.

A prepared demonstration export file with fictional data. This illustrates the format; fields and structure are adapted to your software. The workbook is in Lithuanian.

Application and export examples with demonstration data.

How we start

We start with your sample documents: agree the fields, verify extraction accuracy and prepare exports to match your software requirements.

02 / DI In production Document AI · data extraction
PythonLLMPDFTestaiCloudflare

Local document processing with DGX Spark

Local AI model and document processing infrastructure
Problem

Invoice data had to be typed in from PDFs by hand. It took time and left room for mistakes.

Solution

The system reads the supplier, date, totals, VAT and line items, and routes unclear cases to a person for review.

Integration

Validated invoice data can be exported as CSV, XLSX or XML and passed to the client's accounting or other internal system. I adapt the integration to the client's existing workflow.

Demonstration with a real phone photo

A fictional invoice was printed, crumpled and photographed with a phone. The system extracted the fields and two line items, reconciled the totals and returned a verdict.

18,530-document flow (separate project)status as of 2026-09-01
93 % processed without human intervention; not a field accuracy measure 7 % human review or duplicates

Local NVIDIA DGX Spark model: document content is not sent to an external AI service. Deeper analysis takes a few minutes.

Outcome

The integration is live: scanned invoices are handed over to the Pragma accounting system as XML. Verified against real invoices — the suggested item code, ledger scheme, warehouse and VAT code matched what the accountant had entered by hand.

03 / ML In production GPU / ML · computer vision
MLGPUEmbeddingsOCRDocker

Photo restoration & semantic search

Finding and restoring old photos
Problem

Large legacy photo archives: degraded quality, impossible to find anything by content.

Solution

The platform lets people upload a photo, improve it and download the result. Archive photos can be found not only by text but also by what is visible in them, including in Lithuanian.

Outcome

12,000+ photos in the library, +12% more text found after OCR reindexing, search by what is visible, in Lithuanian too. Restoration raises resolution 4× in ~2 min, with an automatic quality gate.

04 / OPT In production Optimization · full-stack web
PythonFlaskILPPuLP/CBCSQL

Cutting-plan optimizer

Less manual planning and waste
Problem

Film cutting plans were calculated with a calculator and entered by hand into an old, unsupported system.

Solution

I built a tool that chooses a cutting plan, helps manage materials and prepares reports. I migrated the full history, so users kept it and the old insecure system could be retired.

Cutting sheets for production

The system prepares clear production sheets for different numbers of lanes. Both examples use demonstration data only.

Outcome

19,600+ historical layouts migrated to the new database and production is run from it. Honestly: a backtest over the whole history showed manual planning was already ~97% efficient and only 0.4% of layouts could be cut better, so the real value is reliability and error catching, not saved material.

05 / OPS In production Infrastructure · DevOps
ProxmoxCephLinuxZabbixFRR/OSPF

High-availability cluster & monitoring

A reliable base for working systems
Problem

Everything above needed a reliable, self-managed foundation to run on.

Solution

The three-node infrastructure keeps data in three copies, has redundant 10G networking, backups and monitoring. If one server fails, virtual machines can move and keep running.

Outcome

Silent node freezes are handled: a hardware watchdog reboots the node in ~10 s and the virtual machines restart on the remaining two — tested by actually taking one server down. Critical events arrive by email and backups are incremental (6.8× deduplication).

06 / IOT In production Production workstation · scales · labels
PythonTkinterSerial/COMPDFPyInstaller

Weighing & label-printing workstation

Weighing, labels and records in one place
Problem

An operator needed one reliable workstation to read a stable weight from the scale, generate a label and leave an auditable weighing record.

Solution

The application reads a stable weight from the scale, prints a label and saves a record. After a restart, each machine returns to its own state, and every weighing can be tied to an order, batch and pallet.

A short demonstration from the workstation

The application is used beside production equipment: the scale, operator screen and label printer are all at one workstation.

The application is deployed at the operator workstation and runs without a separate Python installation.

Outcome

The program runs on the production PC and a new version is installed by copying a single file. 6 separate machine contexts, scale connection in 2 s, 93 automated tests before every install.

Profile

About

Paulius Sutkus

I'm an engineer who cares about finishing the job: understanding the problem, building a solution and making it easy to use. AI matters to me when it truly saves people time or helps them avoid mistakes.

I start with a concrete work process, but see the wider system: how it connects with data, people and other tools. I work from the first pilot through integration and maintenance: using AI, Python, databases, web interfaces and Linux infrastructure.

I work on B2B projects, contract engagements and permanent roles that need ownership from process analysis through to a working integration.

“Stability, precision and planning.”

That is how colleagues describe my work.
AI / MLLLM · embeddings · GPU · OCR
LanguagesPython · SQL · Bash · PowerShell
Web / back-endFlask · REST · SQLite/PostgreSQL
OptimizationILP · PuLP / CBC
InfrastructureProxmox · Ceph · Linux · Docker
OpsZabbix · FRR/OSPF · HA · backups
On the side

I build and maintain websites for small businesses.

Let's talk

If you're looking for someone who can take an idea and turn it into a working tool, get in touch. I'd be happy to talk through any of these projects live.

LinkedIn

Available for: B2B projects · permanent roles · contract work