I see a problem, I build the fix.
Most of my work starts with something that annoys me. A test suite that cries wolf every morning, or a language model that invents buttons that do not exist. I build the thing that makes it go away, and then I measure whether it actually did.
My job title says test automation. I ended up there by accident, and it turned out to be a good place to build things, because there are real systems to break and a reason to try new technology on them. These days that mostly means AI agents and the tooling around them.
now: running the thesis experimentlast commit: Ironman 70.3: 231 days
Work
- 2026Agents that argue about architectureMy master's thesis asks whether LLM agents that debate a software architecture judge it better than one model on its own.research
- 2026Teaching an LLM where the buttons areA crawler maps a web app so that AI-generated tests stop inventing selectors. Invented selectors went from 18 % to 0 %.client work
- 2026flakeproofWhen CI goes red, flakeproof tells you whether the app broke or just the test.open source
- 2026Where in the code is this bug?A benchmark on 918 real GitHub issues that measures how well different search methods find the file that needs fixing.research
- 2026cfd.houseI put rubber ducks and race cars into a virtual wind tunnel every night and post the airflow with a house track.side project
- 2025prompt-contractsContract tests for LLM prompts, so a model update cannot quietly break an integration.open source
- 2026Sorting tickets nobody wants to sortA classifier sorts incoming Jira tickets, a person confirms, and every correction makes the next round better.client work
Smaller things
- qrdrop
- AirDrop for everyone. Drop a file, scan a QR code, it goes peer to peer over WebRTC.
- Leasing documents, compared
- Automated target-actual comparison of leasing documents, rerun on every system update. At work.
- Tender matching
- An LLM rates freelance tenders by how well they fit, with a dashboard. At work.
- Time tracking sync
- Copies booked hours from Jira Tempo into Timebutler every night, running on my own server.
- Driverless race car
- Motion controller for the autonomous car of DHBW Engineering Stuttgart, 2023 to 2024.
Day job
Since 2022 I have been at TestGilde, a software testing consultancy near Stuttgart, first as a dual student and now as an engineer. On client projects in automotive and e-commerce I automate end-to-end tests, for example for the online vehicle sales of a car maker, and I lead test management for a service contract platform and its rollouts across Europe. In between I build our internal AI projects, which is where several things on this page started. Clients stay unnamed here.
Research
I am doing my master's in Digital Business Engineering at Reutlingen University. The thesis tests whether LLM agents that debate an architecture evaluate it better than one model on its own, when both get the same budget. The experiment runs this autumn, submission is in early 2027.
A footnote
In fourth grade my teacher recommended the Realschule, the middle track of German secondary school. Neither of my parents went to university. I did a dual bachelor's degree anyway and I am now writing a master's thesis. I do not take an answer as final until I have tried it myself.
Freelance
I take on a few projects next to my job. Remote by preference, on site in the Stuttgart region if it helps, in German or English.
Good fits are AI systems that have to work reliably, automating something that currently eats hours, test automation people actually trust, and the occasional website or internal tool. I am the wrong person for work that is mostly paperwork, or for projects that need two months of access requests before anyone writes code.
How I work: concept first, then a prototype you can click within days. Before anything goes live there are numbers that show whether it does what it should.
Write to me with a few lines about the problem: philipp.melikidis@gmail.com
Outside work
Outside work I am training for an Ironman 70.3 in May, so most evenings end in a pool or on a bike. I DJ hard groove, house and trance and make my own beats, which is also why my wind tunnel videos come with house. And a ThinkPad in my living room runs the simulations at night and a few services for me during the day.