Experience
I’ve worked here and there with a general theme of teaching and banking, two things I never exactly dreamed about, but I seem to really enjoy.
At Íslandsbanki I get a Teams message every week about some new project that they think me capable of building. I scope the process, contact the necessary people, and get to work. It's a loop I haven't got tired of yet and don't imagine getting bored of any time soon. I've built some plugins, skills, and Power Apps that people seem to like. If they're happy, I'm happy. It's a great team that I can talk about films with. What more could a man ask for?
- Building and maintaining automation workflows to streamline internal banking operations.
- Building Copilot plugins and skills for internal use across the bank.
- Improving and optimising existing Blue Prism RPA processes that were originally built without performance in mind.
I got my first look at corporate work at Arion Banki, where I was given the task of “figure out how to test our agents”. The current process was manual, ad-hoc, and didn’t scale at all. It was both terrible and great to realise that Steve Jobs was right: everything around me was made up by people, and now I was one of them. After using Google ADK to sort that out, the team held a meeting about AI safety. Since I was the only one who had played Lakera’s Gandalf game, where you trick an increasingly difficult AI model (of course disguised as Gandalf) into giving you a password, I was deemed the most qualified on the team to add safety tools to the user-facing agent.
- Designed and built an automated evaluation framework for AI agents, replacing a manual testing process and making agent quality measurable and repeatable.
- Created an evaluation dataset from the company website to benchmark agent performance against realistic queries.
- Implemented input and output guardrails to improve the reliability and safety of the AI system.
- Debugged framework and integration issues through source-code analysis and Azure LLM integrations.
- Worked in an agile team using Git, Jira, and code reviews, shipping changes into a production environment with strict security requirements.
I applied to be a Teaching assistant at Reykjavik University as soon as I could be one and TA'd at least one course every semester I could. I was quite popular as a TA as I always had many immigrants in my classroom (that was what we called students who left other TA sessions to come to my sessions). I always heard that I was the most enthusiastic and passionate TA around, which I appreciated.
- TA for Algorithms, Calculus, and Discrete Mathematics I & II in the Computer Science programme.
- Led weekly sessions for ~60 students, explaining algorithms, data structures, and mathematical concepts.
- Strengthened my ability to communicate complex technical concepts clearly to students at a range of skill levels.
Funded by a Rannís research grant (Nýsköpunarsjóður námsmanna), Teitur Hrólfsson and I carried out a study at Reykjavík University. The project ended up being called “MELD: A Multimodal Ensemble of Lightweight 1D Convolutional Neural Networks for Sleep Stage Classification”. We explored whether evolutionary search could find better CNN architectures for sleep staging. Early on I thought of a rather clever name, SLEAP (Sleep Labelling using Evolutionary Algorithms and Pytorch); this name fit for most of the project, but that was before realising a horrible truth: the bigger the CNN architecture, the better the model. I feel like I stumbled upon the same truth that all of the AI companies had realised some time earlier. In the study, I focused on tuning kernel sizes via a genetic algorithm; it turned out that kernel sizes alone moved the needle very little, which was a useful (if humbling) result in itself.
- Implemented genetic algorithms in Python (PyTorch) to search CNN architectures for sleep-stage classification.
- Built training pipelines for multi-hour experiments across sleep datasets.
- Iterated on optimisation strategies through systematic experimentation.
- Co-authored a full technical report for Rannís documenting the methodology, experiments, and findings.
In 2020, I was talking with my coworker at Kópavogur Library and he asked me if I had done any programming, since I “looked technical”. I said no to him but he somehow convinced me I was qualified enough to teach others. He told me that as long as I knew a little bit more than my students I could teach them. This, to be fair, still remains my teaching philosophy, as long as you're an interesting enough speaker. We brought iPads to libraries so kids could code hands-on, with the goal of making tech and programming accessible to anyone no matter their background or whether they had a device of their own.
- Introduced absolute beginners to programming fundamentals using Python.
- Taught programming to children in free, drop-in workshops hosted at local libraries.
- Brought iPads so kids could get hands-on coding experience without needing their own devices.
- Helped make tech education accessible to everyone by removing the usual barriers to entry.