We have new training videos, exercises, and certifications for Vaadin 25. What took us so long? I'm glad you asked.
Why updating training content is hard
Updating training materials to a new major Vaadin version has always been a big task. Ideally, when a new release ships, we'd be ready to train people on it right away. But it's never been that simple. The features that make it into a specific major release aren't usually locked in until the last possible minute. That makes sense: we want to bring in as many new features as possible, but only those that pass the quality gateways. As a consequence, it's hard to know when to start updating the training content, and the people responsible for it (including yours truly) are often busy, or at least good at finding other equally-important-but-less-annoying things to do.
Our training Product Owner, Tatu, and I have been discussing ways to improve this for years. Filming videos has been the biggest bottleneck: it requires finished materials, willing people, and everyone's schedules aligning. Performing on video isn't always easy or appealing for us Elder Millennials and Gen-Xers. So the question was, could we automate it?
Building an AI-assisted video production pipeline
We checked whether an off-the-shelf solution would work first. The SaaS video generation market is surprisingly diverse, but nothing we looked at was the right fit. Either the quality wasn't good enough, or the tooling didn't match our intended workflow. So we decided to build it ourselves. How hard could it be?
The question everyone has at this point: did you use AI to generate the videos? Well, yes and no.
We wanted a reusable process where a training module's script and its exercises are the source of truth, and everything else derives from them. The training script generates both the slides and the narration. Combine those two things and you can generate a video and up-to-date certification questions. A few API calls, some libraries, some glue code, and Claude was happy to help with that part.
From Vaadin 24 scripts to Vaadin 25 videos, step by step
Of course it wasn't quite that simple. How did we get the training scripts in the first place? Using an LLM to write them didn't work (as expected), but we already had existing Vaadin 24 videos and matching slide sets. From those, we extracted the slide structure and used ElevenLabs' Speech-To-Text API to create initial script drafts. Those drafts were rough, but at least we had a starting point. We then ran a cleanup pass with OpenAI's lighter models to fix common mistakes. "Vaadin" was not easy for the STT engine to get right. The Vaadin 24 exercises and code examples also needed upgrading to Vaadin 25, and here again Claude was helpful, once we convinced it that "it compiles" is not sufficient proof that the migration worked.
For generating slides from scripts, we wanted consistency and repeatability. Claude or Gemini would happily generate something that looks OK most of the time, but it turned out there was no need to involve an LLM at all. We created a master template slide deck with slotted layouts based on the older manually-crafted decks. The script just specifies which layout to use, and the process drops the content into the right slots. Just like web components.
To produce a video, a small program plays through the generated slides and applies a voice-over from a Text-To-Speech service. We went with ElevenLabs' "Shaun" for production, and OpenAI's free Edge TTS during development. The same process gives us captions for free. Previously we used captioning services where results still needed manual correction every time. Voice generation is now the only "AI" step in the whole pipeline, which is what I meant when I said we only kind of used AI.
Here is the simplified overview of the process:
Was it worth it?
The whole process involved 7 people working in and out, and so, so many rounds of code and content review. We even built ad-hoc Vaadin UIs just to make it easier to read through all the text.
Was it faster than doing the Vaadin 25 update manually? Hard to say, probably not. Was it worth the effort? Yes, because now we can fix issues in slides and videos almost instantly. New branding guidelines? Just update the template deck. New components in a minor release? We can add them without starting from scratch. Whole new training tracks are now a real option instead of a no-go.
Go check out the Vaadin 25 training videos and get certified. Content correctness is most important, but so is the delivery. If you prefer a real human instructor over the AI voice, we still do traditional Vaadin trainings and workshops too, online and in person.