Simplify
Remove unnecessary steps, decisions and handoffs.
WE IMPROVE OPERATIONS AND DECISION FLOWS
Things canwork better.
If a process depends on copy-pasting data, manual checks, email chains, spreadsheets and workarounds, we first find where time and money are actually leaking. Only then do we choose the simplest solution that removes the friction.
BeforeGoal: less friction
Illustrative status quo
E-mail
Excel
ERP
Check
Approval
Report
manual handoffsduplicate datawaitinghard to scale
After simplification
Source
Flow / rules
Outcome
Fewer steps.We simplify first. Automation is only the next question.
MELIOK — one idea: to better what already exists.
Technology is a tool. The outcome should be a better process.
01 / COST OF THE STATUS QUO
The most expensive process often looks “good enough”.
The problem does not have to look like a failure. More often it is an hour here, fifteen minutes there, copying between systems, manual checks and knowledge locked in one person’s head. At scale, these small frictions become operating cost.
01The same data is re-entered across tools.
02Specialists spend time on work that does not need to be manual.
03Errors return as rework, claims or additional checks.
04Growth requires proportionally more operational effort.
Estimate the cost of one manual process.
We do not invent ROI. Enter your own numbers and see the scale of the problem before discussing a solution.
Work hours / year—on this one process
Cost / year—based on your inputs
The calculator does not promise savings. It only shows the cost of the current way of working based on your inputs.
02 / WHEN NOT TO BUILD
Not every problem needs AI. Not every process needs custom software.
If a process change, a spreadsheet or an off-the-shelf integration solves the problem, that is what we recommend. Custom software only makes sense when simpler options cannot deliver the goal.
Use an existing tool
If a good solution already exists, there is no reason to rebuild it.
Integrate
APIs, synchronization and automated data flow often solve most of the problem.
Automate
Workflows, RPA or backend services take over repetitive rules and manual operations.
Build / use AI
Custom software, agents, vision or models only when they create a clear advantage in the process.
✓
MELIOK rule: we choose the smallest solution that achieves the goal — not the biggest project that can be sold.
03 / THE REAL GOAL
“We need AI” is usually not the business goal.
We ask what should change in the way the company works. Technology only appears as the mechanism that creates that outcome.
Less manual work
Specialists should not spend their day copying, cleaning and repeatedly checking data.
Possible mechanism: workflow / integration / automation
Shorter lead time
The process should move from input to decision without waiting for manual handoffs between people and systems.
Possible mechanism: API / rules / orchestration
Fewer errors
Repeatable criteria, data validation and automated checks reduce the cost of rework and inconsistencies.
Possible mechanism: validation / vision / rules
More control
The process outcome should be visible, measurable and traceable — not hidden in inboxes and local files.
Possible mechanism: app / audit trail / dashboard
04 / HOW WE WORK
Decision process before technical process.
Each step is designed to reduce the risk of a bad investment: from understanding the problem, through choosing the simpler route, to measuring the result.
Map the process
Who does what, in which system, with what data, and where waiting, manual work or risk appears.
Quantify the cost
Time, people involved, errors, delays, additional checks and the cost of scaling the current way of working.
Simplify
We remove steps that do not create value. We do not automate chaos just because it is possible.
Choose the solution
Off-the-shelf tool, integration, automation, custom application or AI — in the order that makes sense.
Build a small proof
We validate the key assumption first. Only then do we expand the solution and connect it to operations.
Measure the outcome
Did process time fall? Are there fewer manual steps? Is the outcome more predictable? That is the benchmark.
05 / SELECTED WORK
We did not “deploy AI”. We changed how the work gets done.
Each case study answers five questions: what happened before, what hurt, what changed, what the new process looks like and why it matters to the organisation.
Document AISensitive dataNDA-safe
Tens of hours of legal work recovered every week.
Working with large volumes of legal documents meant manually preparing content, protecting sensitive data and repeatedly checking material before it could be used downstream.
Situation
Documents contain PII, while still needing to preserve the structure and context required for downstream work.
Problem
Manual anonymisation and preparation is slow, prone to omissions and difficult to scale.
Change
We designed a pipeline that detects and pseudonymises sensitive data locally before content reaches the LLM, then verifies the output and, when needed, restores the data locally.
New process
Ingest → pseudonymisation → model processing → verification → human-in-the-loop for uncertain cases → output with a complete audit trail.
Why it matters
The result is not a “chat with a document”, but a repeatable process that recovers tens of hours of work each week while keeping control over sensitive data.
How it works technically +
FastAPI · PostgreSQL · token vault · Regex/NER · LLM · verifier · reidentifier · asynchronous worker · audit log · React/Vite · Docker.
The visual is an original architecture reconstruction. It does not show client documents or data.
Computer VisionManufacturingNDA
17 lines. One digital quality standard.
Across multiple production lines, OK/NOK detection alone is not enough. The result must be repeatable, recorded and comparable across locations, shifts and time.
Situation
Quality control covered 17 production lines and required a consistent way to record the result.
Problem
Without one digital standard, it is harder to compare results, trace deviations and build data for process improvement.
Change
The computer-vision layer turns image observations into a standardised quality result that can be stored and used downstream in the quality process.
New process
Image → detection → OK/NOK result → record → traceability → analysis over time.
Why it matters
The value does not end with automated detection. It creates a common control standard across 17 lines and data that can be analysed and improved over time.
Due to NDA restrictions, we do not publish workstation photos, production data or the client interface. The visual below is an abstract representation of the inspection architecture.
Forensic softwareOpen sourceproject contribution
An expert tool for trace comparison, annotation and documentation.
Contribution to Forensic Biometrics Studio — a desktop application for side-by-side forensic image comparison, marking corresponding features and creating structured comparison material.
Situation
An expert works with two images at once and must preserve annotation precision while panning and zooming.
Problem
New annotation types require consistent drawing, mouse handling, visibility, translations, shortcuts, and save/load behaviour.
Contribution
Extending annotation classes and handlers, drawing logic that accounts for viewport positioning during zoom, and integration with the toolbar, shortcuts and annotation serialisation.
New workflow
Load images → mark corresponding features → compare → save data → generate structured material for further work.
Why it matters
Public project on GitHub ↗
Software should support the expert’s workflow rather than force the expert to work around interface limitations. The value is a precise tool and a repeatable analysis workflow.
R&D / ENGINEERING
Projects that show range — without pretending to be commercial deployments.
We separate R&D from client case studies. This section shows the ability to build and test solutions where there is no ready-made answer.
Change the garment. Do not change the person.
A virtual try-on pipeline designed around preserving the person’s identity and the input image: garment classification, clothing parsing, generation only inside the mask and a hard freeze outside it.
From measurement to physical action.
A prototype combines temperature, humidity, light and soil-moisture sensors with local logic controlling doors, irrigation and feed dosing. Data is sent to InfluxDB while decisions can be executed locally.
06 / WHERE TO LOOK FOR LEVERAGE
Processes where technology most often makes sense.
You do not need to know what solution you need. It is enough to recognise one of these kinds of friction.
Too much reading and re-entry
Extraction, classification, validation, document generation, workflow and sensitive data.
Data moves between spreadsheets, systems and email
Reporting, controlling, document matching, statuses, rules and controls.
The process grows with headcount
Workflows, integrations, repetitive decisions, recurring tasks and data handoffs.
Inspection needs consistency and traceability
Computer vision, sensors, automation, traceability and engineering tools.
A person is acting as the API between systems
Data synchronization, APIs, ERP/CRM, automated import/export and removal of copy-paste.
The information exists but is hard to find
Search, RAG, classification, structured knowledge repositories and decision-support tools.
The spreadsheet has outgrown itself
Custom tools, forms, dashboards, roles, audit trails and integrations with the existing environment.
A calculation or decision happens outside the right tool
CAD/Python tools, calculation automation and logic embedded where the engineer actually works.
07 / REDUCING DECISION RISK
What happens after you contact us?
We do not start by quoting an “AI implementation”. We first check whether the problem is worth building a solution for at all.
1You briefly describe the process and desired outcome.
2We frame the questions about scale, cost and constraints.
3In the call, we identify the friction and simpler alternatives.
4If the project does not make sense, we say so. If it does, we propose the smallest sensible next step.
PROCESS DISCOVERY
One process. 30–45 minutes of focused analysis.
100 PLNProcess and bottleneck analysis, 2–3 improvement directions, assessment of where automation makes sense and a recommendation for the next step. If we proceed to a project, PLN 150 is credited against the first invoice.
Payment via Stripe. Choose a time after payment.START WITH THE PROCESS
Show us a process that should work better.
You do not need to know whether you need AI, an integration or an application. Describe the current way of working and the outcome that would matter.
01We reply by email and clarify the context.
02If there is potential, we schedule a process call.
03If a simpler option is enough, we recommend it instead of a project.