Web Summer Camp 2025: behind the scenes

In July 2025 I had my first public speaking engagement: I held an LLM serving workshop at Web Summer Camp 2025.

Web Summer Camp by Netgen is a long-lived web development conference primarily focused on things that power the organization company, Netgen: PHP (Symphony), JS, UX and project management. There’s a twist, though: most of the presented content aren’t just talks, they’re practical workshops intended to give the best experience to the audience. Since I’m probably as far as it gets from web development, it’s an interesting story how I ended up there.

The year before my workshop, Ivo Lukač, founder of Netgen, approached me and shared with me that they plan to introduce a new track the next year, focused on Python and AI, and whether I’d like to apply to hold a workshop there. I always dreamed about having a public speaking career, and incidentally, I had an idea and some materials about a talk i wanted to hold once re model serving. So i said: “Sure, I’ll think of something”, wrote a description and sent a submission. One day, I received the acceptance mail, and the rest is history.

A more interesting story, however, is how the workshop I held actually came to be.

Everything old becomes new again

I had conceived the original idea early in the history of “MLOps” how it used to be called, back in 2022. I was supposed to present traning and serving a dog/cat classifier. There weren’t a lot of options back then, so I had prepared a simple TorchServe script and a Triton inference server example. That was also a different time, when we (as Deep Learning industry) had plenty of architectures to choose from, each specialized for its usecase. Upon returning to that repo, I learned that almost everything had changed

First, there are plenty more serving options now: Triton is still here, there’s this new (to me) kid on the block called vLLM, and there are dozens of local model serving tools. Second, everything are LLMs now, even for multiple modalities. I took the simple decision and stayed with what I knew the best (Triton), and found a new use case that leveraged LLMs.

At first, I prepared the workshop with Gemma 2 in mind. I had both parts of my workshop more or less ready, when Gemma 3 came out, which showed much better results. Unfortunately, the serving technology lagged quite a bit behind and Triton simply did not support that new architecture. Due to two poorly timed releases and bugfixes on the Triton inference server repo, I had to wait until early May to actually start working on the second part of the workshop. Fun times

Also, with a switch to Gemma 3, I had to give up multiple modalities because, simply put, Triton Inference server still does not support image input Gemma v3. I had to scrap my math problem solver for a simple text-only chat experience.

And to top it all off, I got sick for 3 days in June, just when I was supposed to start developing the workshop. I suppose it was all the stress. But, I did manage to finish and test everything out by the end of June.

Cloud GPUs and where to find them

A whole another show was with the cloud providers. I had (what I percieve as) simple requirements: I want the ability to create ~30 cloud instances, with GPUs (I don’t care which one, as long as it can serve LLMs), on which I could run Docker containers. A secondary goal was not to go bankrupt, but honestly, I was perfectly fine paying a bit more just to give my audience the experience they’ve already paid for by being at the workshop.

This is an excerpt from my notes (the interesting ones):

Provider	why no
Dataoorts	Can’t run Docker
Paperspace	I couldn’t get an approval for GPUs
Lambda labs	They don’t support debit cards??
puzl.cloud	They couldn’t support 30 GPUs
runpod	Can’t run Docker
GCP	Appearance of impropriety

Eventually, I settled with Exoscale. They allowed me to create a machine image with all the drivers, spawn up to 30 instances of said machine image with an adequate GPU, and the GPU clearance was “pay us upfront”. Huge shout out to Benoit and Iva from the support!

Tech Enthusiast Playground Paradise

This workshop was actually an excellent opportunity for me to finally explore some tools I’ve been meaning to explore:

• I finally tried out reveal.js for slides in markdown. I loved it, but then I had the need for more advanced functionality so I switched to their pro web app slides.com. I’ll probably continue on being their user, the site is nice, the functionality is nice (they’re really focused on developers).
• I finally tried out just for command running. I did not used any of it advanced functionality, but it worked pretty okay. By the end I kinda felt more comfortable to try something more complex. I’ll also probably use it more in the future.
• I also finally got the opportunity to work with Packer for creating instance templates. This one decreased instance start time by 80% (~10s of minutes -> minutes).
• astral-sh/uv for Python package management. not my first time using it, but definitely the most serious attempt. I fell in love with it once again, because it Just Works™. For the workshop, I had a virtual instance that ran JupyterLab as a systemd service on startup. I only had to write uv run for the exec and it handled all the env setup (dependency installation) for me.

jj vcs

The most useful tool I worked with during the creation of this workshop was definitely the jj VCS. I’ve been actually running jj for private projects for over a year now, so it can hardly be called exploring, but I finally had a great use case for some of the more complex workflows:

• Frequent creation of new changes just to try something new out
• jj mega merge + jj absorb. Mega merge is when you merge many changes (10+) into a new child change. You can then make edits in that merged child as needed (making stuff build, syntax, typos), and jj absorb distributes the edits back to the parent changes accordingly. Probably the closest thing to magic I felt in years.
• moving changes in files to an earlier point in history. For example, quite late in the workshop development process, I realised that I needed another dependency. A simple uv add in the middle of the change chain, and jj squash to move the lock file to the first commit of the workshop branch. The rest of the chain gets rebased automatically. Literal chef’s kiss.
• Rewriting history in general, such as splitting changes, changing the order of changes, their descriptions, and rebasing all the descendants automatically. In order to have a clean sequence of changes with git, you basically have to one-shot it, as it’s pretty hard to retroactively fix mistakes. And I really cared about having every change be a self contained unit (see learned lesson #2 below).
• I finally conceded and started using a visual conflict resolution tool (Meld). I’d say it’s better than vimdiff, but also, the difference in experience is not that big. I might go back to vimdiff, or I might try something else (kdiff3?)

All in all, I understand that the use case of 1) a single developer 2) working on a fixed functionality in a single repo 3) and that functionality is an ordered sequence of changes that primarily tell a story, is not the most representative use case for a VCS. But I truly mean it when I say that this tool enabled the workshop to be as complex as it was. Without jj the scope would be much smaller. If you haven’t already, I recommend you try it.

Lessons learned

1. Content creation is hard business

From the moment I started working on the workshop until I actually held it, >6months of wall time have passed. Of course I didn’t work during this entire period, but I did work for a larger chunk than anticipated. I’d estimate that if I worked on the workshop for 10h a day, 6 days per week, I’d get it done within a ~month. This is a huge amount, especially with a full time job and with other personal commitments.

1.b Exploration is a huge part of it

If I knew exactly what I wanted to create, I’d be done in 3 days. This way, 80-90 percent of time actually went into exploring the landscape. Which model to try out? What use case? Which model serving technology to explore? How deep do I go with the fundamentals? All sorts of things had to be answered during the production, multiple times, with an expensive evaluation of answers and frequent misses. One thing in particular that I spent a lot of time on was installing and aligning GPU types, cuda drivers and dependencies. I tried out 5 GPUs with Exoscale alone, only to find out it’s too old and it does not support something I needed. Truly an unfortunate state of affairs.

2. Apparently I’m a perfectionist now??

I was always proud of myself to be the person that knew when enough is enough, how not to let the “perfect” be enemy of the “good”, and when to stop investing effort. Also, I’ve never really saw myself as a detail-oriented person¹, I’m definitely more of a knowledge-breadth guy than a knowledge-depth guy. But working on this workshop woke a pedantic side of me that I never experienced before. On my first pass through the entire thing, I ended up with a full page worth of TODOs in my Obsidian. And by pure magic, after the second pass to solve some of the entries, the list got even bigger. I’m not trying to do a public introspection session here, but it’s probably because I cared about this workahop way more than the other things I did before. This realization also sort of scares me, because it implies I’m not able to do Good Work™ without being personally invested, and caring about something is a constrained resource. Let’s see if I can hack myself around this.

3. I love public speaking

There’s something weird in my brain. The moment I step on the stage, it immediately goes into a new, higher functioning state. It’s like switching to a very high gear. The brain starts working really quickly, my creativity increases, I’m so excited, and I truly feel like home, like I belong in front of people telling stories and making dumb contextual jokes³. Sadly, I’m completely alone in this in my circle of friends, so I can’t relate with anyone close. But it’s true: I love public speaking and I feel at home on the stage.

Open questions

AI usage

Here’s a confession: I hadn’t use any sort of AI for developing this workshop² up until the last day where I could not be bothered to write, again, a Docker Compose file, which I’ve done more times than I can count. There are plenty of reasons for this, and not all of which are practical, but the most important one was that this workshop was, primarily, a way to express myself. The emphasis is on me, a human, that likes exploring stuff, creating things, and sharing knowledge. Creating this was genuinely a fun experience which I’m likely to repeat in some form or another.

That said… I really spent a lot of time on this, and I sacrificed some other parts of my life because I wanted everything to be accessible and clear and nicely presented. Here’s the kicker: those are conflicting goals. The way I like telling stories might not be the best way to explain a thing. And this seems to be the promise of AI, to be better at something than the humans (+ obvious time saving). The problem is, now after creating the workshop and reflecting on what I created, I can’t really think of a way where an LLM would’ve helped me. And that is actually what I want answered: is there a use out of LLMs for me developing workshops?

---

Time will tell… … and more opportunities will help me! If you’re reading this, and would like me to host a talk somewhere, please let me know