Language Models
Finnish and multilingual LLM evaluation, benchmark tasks, datasets, Hugging Face tooling, lm-evaluation-harness, and research workflows that involve both code and patience.
The scenic route through language, machines, and their occasional bad habits.
About
I am an ICT engineer who wandered through copywriting, telecommunications, Linux servers, cybersecurity, and finally the strange weather system of large language models. The short version is this: I like systems that have language in them, and language that has systems underneath it.
My current center of gravity is Finnish and multilingual NLP. With the TurkuNLP research group, I have worked on large language model evaluation, benchmark construction, corpus processing, and the unglamorous but important craft of making data less embarrassing before a model eats it.
Cybersecurity is the other half of the contraption. I study it formally, practice it through Linux and network infrastructure, and keep returning to the question of what happens when AI systems become part of real security workflows. It is not enough that a model sounds clever. It should also fail in interestingly small ways.
Before all this, I wrote words for money. That older trade has not vanished; it now sits quietly behind the keyboard, helping me explain technical things to humans who have not agreed to suffer unnecessarily.
Working Shape
Security
System and application security, network security, firewall and IPS thinking, ethical hacking coursework, and an interest in AI systems that can be inspected without squinting too hard.
Systems
Linux servers, Bash, Docker, Proxmox, VPNs, monitoring, self-hosted services, and enough homelab practice to respect both documentation and cable labels.
Communication
Technical writing, Finnish language awareness, research reporting, copywriting muscle memory, and the small mercy of saying complicated things plainly.
Public Traces
Some work leaves a paper trail, which is convenient because otherwise research would just be a mist of terminal output and meeting notes.
An Expanded Massive Multilingual Dataset for High-Performance Language Technologies (HPLT)
Co-authored ACL 2025 paper on HPLT v2, a large multilingual data collection for language technologies.
arXivFIN-bench-v2
Co-authored benchmark suite for evaluating Finnish large language models across consistently formatted tasks.
arXivHow I Tend To Work
I am happiest with problems that begin as fog and slowly become a checklist. Research, security, data processing, and infrastructure all share that small pleasure: first you are confused, then you are less confused, then you write down what broke.
The useful part of my background is not that it is tidy. It is that I can move between disciplines without pretending they are the same thing: linguistics is not DevOps, cybersecurity is not copywriting, and somehow a good Bash script may still need all four.