Software development & system integration
I build applications and automation tools that connect APIs, data and AI services. I'm in my final semester of Electronics and Telecommunications at Gdańsk University of Technology, and I've been into automation since primary school.
In short
I build automation tools, desktop applications and interfaces for technical systems. I take each one myself from the idea to a working version.
I use a few of them every day, such as Yapper and my own email tools, so I quickly see what breaks.
See the projects
Core strengths
hardware
I have done lab measurements, circuit simulation and PCB design, so writing software for devices, embedded systems or IoT would be a short step for me.
LTspice / Micro-Cap / KiCad / EAGLE / Vivado / Basys 3
systems
In projects I work with Windows and Windows Server, virtual machines, Linux and Docker. Small repetitive tasks go to AutoHotkey scripts.
Git / Linux / Docker / Windows Server / VMware / AutoHotkey
software
I write tools, API integrations and data processing in Python and JavaScript. I know C/C++ and Java from classes, projects and learning on my own.
Python / JavaScript / REST APIs / C/C++ / Java
Education and experience
2023 - now
Electronics and Telecommunications, Computer Electronic Systems
Lead software developer of the SeeSky radio telescope in the SimLE research club: backend, tracking interface and an interactive sky map. I am also responsible for programming and testing the SDR receiver. Until the radio telescope is built, I test it on placeholder data.
Final semester ends 2 February 2027
09 - 10.2024
Internship, Telecommunications and Media Division
Set up a Windows Server 2019 environment for certificate handling, built a REST endpoint that passes requests to Microsoft AD CS over DCOM, and wrote the setup guide and documentation.
REST, Windows Server, AD CS
2019 - 2023
Polytechnic profile, mathematics and IT
I took extended maths and computer science here.
My projects
Integration and automation
Collects listings from eight job portals into one application, filters them with shared rules, researches promising offers with a language model and sends notifications. A separate module fills in application forms - I tick the consent boxes myself.
Lead software developer, SimLE
Software for the student research club's 3 m radio telescope. I lead the backend, the tracking web interface and the interactive sky map.
Java and object-oriented design
A Java incremental game about building up chemical-element production, from a coal mine to a reactor making synthetic elements.
Tools and other projects
YapperA lightweight speech-to-text app with two push-to-talk channels and online or offline transcription that copies the text to the clipboard automatically.Python, PyQt6, Groq, VoskDaily use
Turf MatchmakingA Krunker userscript that finds lobbies by configurable filters, checks live data and retries. Other players use it too.JavaScript, APIIn use
Inventory GeneratorA concept generator that builds a Minecraft-style inventory from a searchable item database, with configurable slots and amounts.JavaScript, HTML, CSSLive demo
Password ManagerA password manager that works as a credential vault on Windows and Linux, built as a student team project.Desktop, teamPrototype
BlackBoxA C++ terminal logic game in which the player finds hidden atoms by tracing how rays leave the board.C++University project
Florist websiteAn online business card for a local florist that I built on my own for a client. It went through many rounds of revisions, and the client still uses the font I extended with Polish characters in FontForge.BootstrapLive demoAbout me
I am in the final semester of Electronics and Telecommunications at Gdańsk University of Technology. I build applications and automation tools.
I break a problem down like a tree: one root, then categories, subcategories and the links between them, until every leaf is something I can build and test. I treat my own work the same way, so anything I keep doing over and over becomes a tool.
Outside code I like geography and blending music, I go to the gym, and I play with AI audio-separation tools and architectural visualisation in ComfyUI.
I use Claude Code and Codex every day, and they write a good part of the code. Work from one model gets reviewed by another of similar strength from a different company: a fresh context helps, but a model trained differently also reasons differently and catches other mistakes. When something has to be right, I have a stronger model walk me through the whole file until I understand what it does and why. Everything runs through Codeg: it keeps my Claude Code, Codex and OpenCode subscriptions in one place, and a lead agent hands out tasks based on how capable and how fast each model is.
At a company the data matters most. Sensitive data stays inside a sandbox and, depending on how sensitive it is, does not go to OpenAI or any other model provider at all. Agents work on placeholder data with the same structure, and the real data goes into the real system in bulk only after the change has passed its tests.
Below is the workflow I would use on a team, where a mistake costs more.
Read-only agents study the problem: the existing codebase, its documentation and external solutions such as APIs and libraries. The problem is then decomposed into components, and the plan defines the modules, their interfaces and the files each component may modify. A human approves the plan before implementation begins.
read-only agents + humanThe lead agent splits the plan into independent blocks with defined inputs and outputs. No two blocks edit the same file, so they run in parallel and merge without conflicts. A failed block is redone without touching the rest.
lead agentEach block gets its own coding agent, on its own branch, in a sandbox. The agent can run only the commands its task needs and sees only the files the plan lists for it. An agent that is confused or misled cannot reach beyond its block.
coding agentsA different model of similar strength reviews each diff in a fresh context, checks it against the plan and runs the tests. A failed block goes back to its agent with the review notes, and the loop repeats until the block passes.
reviewer agentThe change runs on placeholder data first, then on real data. It is merged only after a person approves it and can explain how it works. Larger changes get a backup first, so they can be rolled back. Problems users report afterwards go back to G1 as new requirements.
humanThe agents are boxed in mainly because of prompt injection: a leak becomes possible when the same agent reads untrusted content, can reach private data and can send data out (the lethal trifecta). A model rates its own code too highly, so a different model reviews it, and one green test run is not enough (passk, not pass@1). Every new dependency is checked at its source, because models invent package names and attackers register them.
I haven't run agents in a company's production environment yet. Based on my own knowledge, drawing on Building effective agents, Claude Code best practices and the OWASP GenAI Security Project, I would propose a similar setup there.
