MLOps Teaching Assistant and outreach

MLOps Teaching Assistant

My supervisor’s wildly popular three week course on Machine Learning Operations was here in January. This time around it has about 450 course participants. We agreed that it made a lot of sense for me the be a teaching assistant in the course since I have to fulfill the requirement of 150 teaching hours as per the rules of the PhD School at DTU Compute.

I had two main tasks: Firstly, helping out students and secondly developing a Python package that would scrape the groups’ projects and provide feedback in a scalable way. I’ll get back to that second one.

The course is super well-structured, allowing people who’ve never opened the terminal to follow along just fine. You’re provided with step by step instructions for tools such as Docker, experiment trackers and Cloud infrastructure. The students form groups of about 3-5 people and do a project where the assignment is to do a full end-to-end machine learning project, putting together much of what they have learned throughout the exercises.

TA’ing in age of AI

Throughout the three weeks there would always be a few TAs to either be physically present on campus or on Slack where students could ask questions and get help. I had been looking forward to it as I like the material quite a lot and was interested in what people struggle with. But there was not a lot of struggle. At least visible to the TAs. Turns out that LLMs have, in large part, taken the job of the TA as being a patient, helpful resource that’s always on duty. Is that a good thing? I would argue yes, absolutely.

I had some time to think about learning objectives and what makes you good at MLOps. One of the conclusions I made was that stamina, i.e. going at a problem for way longer than you had anticipated with unknown unknowns popping up left and right, and finally reaching a solution, is a key meta-learning from this course. There will always me compatibility issues and problems with dependency management. That is most likely never going to completely disappear. So you learn to deal with them. And I believe that an important learning from this course is to have a clear vision of using a specific tool to solve a problem and going at it until all of the rough edges are dealt with and it finally just works.

That’s why I think, for this course at least, that LLMs as TAs are a good thing. It’s not really all that important that you understand the details of why tool X is incompatible with tool Y but you can fix it by introducing Z. It’s important to learn to spot those patterns and decrease the feeling of helplessness the next time you’re faced with software problems.

Scalable feedback system

Nicki had put together a sort of leaderboard for the previous year’s group projects. It would automatically collect information about every group’s project repository in order to get a few high-level indicators for how the project was going for any particular group. This could be number of lines of code, inclusion of a requirements.txt-file, and so on. He wanted to expand this idea and have an LLM look at the repos as well to provide more qualitative feedback. I thought that it sounded like a fun idea.

It culminated in the rasgaard/mlops_mentor repository on GItHub, where I have collected Nicki’s scripts and structured it so I hope it will be maintainable for next year’s round of the course. The resulting feedback and statistic are saved to a HuggingFace dataset and presented in a Gradio app, also hosted on HuggingFace.

Outreach

One of the things I have been looking forward to with regards to starting a PhD is the networking. Being my supervisors only PhD student and focusing on topics that are not very well represented at the university puts me in kind of an awkward position in terms of collaborations.

However, I don’t think it’s beneficial to focus on the negatives so I take it as an objective in itself to create such a network for myself. That in itself can be a bit socially nerve-racking but I believe that it’s a muscle you build up. Cold e-mailing/LinkedIn connecting/etc. is weird and anxiety inducing but I have only had good experiences so far, making it far less scary to do going forward.

LiteASR

While focusing and reading up on compression of Automatic Speech Recognition (ASR) models I found this paper fantastic. After reaching out to the first author, Keisuke Kamahori, and chatted for a bit we quickly got on a call to talk about other ideas. Now, I’m not expecting that this will lead to anything further but it was just a nice experience. I now feel very comfortable forwarding some early work or ideas his way to get feedback, which he graciously offered.

Pleias

When I saw a post about “Developing Offline-First AI for Community Health Workers in West Africa” I knew that I had to jump in and engage with the post and show my interest. It checks all the boxes for my project and Laerdal’s, in particular Laerdal Global Health’s, focus on healthcare workers in underserved communities with limited connectivity.

I wrote a comment and sent the following message to the Pleias LinkedIn account:

Bonjour! I wanted to follow up on your post about offline-first AI for health workers. I’m an industrial PhD student (started Oct ‘25) at the Technical University of Denmark and Laerdal Medical, focusing on model compression and on-device deployment - specifically use-cases for healthcare workers and simulation.

I’ve followed your work on Bluesky for a while. Big fan of Baguettotron and your ethical approach to training data! Your latest post convinced me that a collaboration could be fantastic!

I am exploring opportunities for an external research stay (a few months in ‘27 or ‘28). I’d love to open a conversation to see if hosting a visiting researcher is something you would consider. I would love to hear your thoughts on this :)

Merci!

reaching the exact 750 character limit. They responded with

Hi Rasmus, thanks for your message. Yes, please write to [email protected].

I of course swiftly sent them an email but have yet to hear back from them. As it’s kind of an out of the blue request I expect that I’ll have to work a bit for a clear response. But that’s for another time.