> whoami
Data scientist and engineer for AI and automation
I build software that takes recurring work off a company's hands. From research all the way into daily operations.
Co-founder
A marketing agency runs on things that repeat: launching campaigns, collecting applications, getting them approved, invoicing, finding what someone wrote down two years ago. I turned that into a platform that handles the repetition and leaves the team with the decisions.
connected through the Meta API. A campaign is built from client data, not from clicking through Ads Manager.
searchable through a hybrid of vector and full text search. Before that, the company's knowledge sat in folders nobody opened.
for 1,544 tests, down from 30 minutes against a cloud database. The suite now runs before every deploy instead of once a week.
Built at Rauch Digital, used by the team and by clients. Not a prototype among them.
From client request to invoice
A campaign passed through eight tools: request form, spreadsheet, Trello, Google Drive, Ads Manager, applicant tracking, accounting. Between every step sat a person carrying data from one place to the next.
One platform a campaign moves through, from the request to the invoice. Copy and creatives are generated from the client interview, the approval page builds itself from those, and once the client signs off, both the campaign in the ad account and the recurring invoice are created.
The CRM where an application does not sit and wait
A campaign brings applications, but nobody gets hired until someone calls back. Applications landed in the client's inbox. Nobody knew who had already been contacted, and an applicant who hears nothing for two days is gone.
Our own CRM for clients. Every application arrives through a webhook and sits as a card in its campaign's list, with a stage the client moves along. Through the WhatsApp Business integration they write from inside the system instead of copying a phone number into another app.
From the ad to the phone call
The application flows lived with an external vendor. Every page was built by hand, measurement and consent were out of our control, and what a page cost was not something we could steer.
Our own funnels on our own domain, assembled from blocks instead of stored as HTML. Draft and published state are separate. An application goes straight into the CRM, and if that path fails, the platform catches it rather than losing it.
From a folder to an answer
The company's knowledge sat in Google Drive, in the project folder and in people's heads. Anyone who wanted to know how a kickoff runs, or what a contract says, asked a colleague or searched for half an hour.
A searchable store with hybrid retrieval: vectors find what was meant, full text search finds what is literally written. Every chunk carries its own context, so a paragraph from the middle of a contract does not arrive without one.
Rauch Digital / Witten, Germany
Responsible for engineering, IT and process automation. I build the platform campaigns are created, approved, launched and billed through.
Ericsson / Stockholm
Six months in house, working on a real analytics tool. It produced a patent and an application that is still running there.
AI Sweden / UNICEF Sweden / Stockholm
A model to identify donors at risk of churning, so UNICEF could follow up where it mattered. Plus a chatbot translating plain language into SQL. Funded by Google.org.
BMW Group / Munich
Cloud computing and agile transformation. Chosen as one of two apprentices for a Udacity nanodegree, Scrum Master (PSM I). Then a QnA chatbot on Azure and Teams that took over onboarding for new apprentices.
Master's thesis 2025 / Aalto University, in collaboration with Ericsson / Stockholm
Conversational Interfaces for Data Analysis: Evaluating Modular Agent ArchitecturesRead it on Aaltodoc
A chat window for people who do not speak SQL. You type a question and get the answer out of the operational data, the analysis alongside it and, where it fits, a chart. I built three architectures and measured them against each other: a simple chain as the baseline, a ReAct agent with retrieval and Python tools, and a modular multi-agent system under a supervisor.
The retrieval-augmented ReAct agent came out ahead throughout, both on correctness and on how well its steps could be followed. The multi-agent system produced clean trajectories but paid for it in coordination latency. Both were evaluated twice: against ten tasks with expert reference trajectories, and in a user study with telecommunications professionals.
work with it today, well beyond the end of the thesis.
built and measured against each other, not merely described.
filed together with my supervisor at Ericsson, and granted.
The work on the agent architectures produced an invention, which I filed together with my supervisor at Ericsson. It was granted.
with Mathew Goonewardena, Ericsson
Diffusion-Based Data Augmentation for Improving the Classification of EEG Signals
Maastricht University
Health data is scarce precisely because it needs protecting. You cannot collect EEG recordings in arbitrary volume, and without volume no model learns reliably. So I generated synthetic EEG signals with a diffusion model to enlarge the dataset and improve emotion recognition.
Transforming the signal into the frequency domain with an FFT raised classification accuracy by up to 20 percentage points over the same dataset without synthetic data. On identical features, though, a GAN came out ahead on three of the four emotion dimensions. Tested on DEAP, EEG recordings from 32 participants.
Twenty repositories on GitHub. Five of them say the most about how I work.
A system that reads NOAA buoy data and forecasts the swell at Huntington Beach. From the scraper to the prediction.
Bachelor's thesis: generating synthetic EEG data with diffusion models to improve emotion classification.
A Claude skill that cleans up code which has grown messy. Readable for humans and cheap for a language model to understand.
With Rebase Energy: weather and wind farm data on one map, so power traders can judge prices better.
Classifying environmental sounds with vision transformers. Sound becomes an image, the image becomes a model.
KTH Stockholm and Aalto Helsinki
Double degree through EIT Digital. First year at KTH in Sweden, second at Aalto in Finland.
Maastricht University
Completed successfully, six semesters taught in English and problem-based. One of them at Reykjavík University in Iceland.
Reykjavík University, Iceland. Exchange semester during the bachelor's.
German (native) · English (C2, studied and worked entirely in English)
My first contact with artificial intelligence came during the apprenticeship at BMW. What set it off was a Udacity nanodegree I was picked for, one of two apprentices.
Until then I had assumed AI was reserved for great minds. For people like Musk or Zuckerberg, not for me. During the nanodegree it clicked: at the end of the day it is mathematics. And mathematics had always come easily at school.
So I tried it and started studying artificial intelligence. First in Maastricht, then in Stockholm and Helsinki. It worked out.
These days the model interests me less than what comes after it: how it becomes something a team uses every day without thinking about it. That is where my work sits.
with my brother Milan
One non-fiction book a week, cut down in a single episode to what actually sticks in daily life. In German, on Spotify, Apple Podcasts and everywhere else.
ListenSummer camp on Ameland 2014 to 2016
KJG St. Karl Borromäus. Two weeks a year, around 50 children, organising and leading the trip.
Swimming instructor 2013 to 2016
Sport-Union-Annen. Four years teaching children between three and ten how to swim.