An introduction to the Workflow Development Kit

[Music] Hey, hey, thanks so much for being here. Um, I'm Renee. I lead the workflow team at Burcell and it's about actually just hit about four months here at Burcell, but this isn't my first time at Purcell. Um, this is a photo of the last time I was here, which was seven years ago. Um I was actually the first intern at Purcell in 2018 and that was my last day

of my internship giving a talk when we launched now 2.0. It was uh basically running Docker in the cloud and that's Tim out there in the in the corner from the next.js team. I ended up leaving uh after my internship and started my own startup and joined YC to do that. So I spent the next five years building a whole bunch of things in between. It started out with a coding agent but this is 2018 so

no LLMs it was hand rolled static code analysis trying to do whatever we could with uh intelligent just code analysis effectively. Um we've actually we pivoted that startup into something else. We started working on data analytics and a uh video generation tool. So don't go too much into my own story, but the sort of the common thread between everything we were building for 5 years

in between always sort of started out as this diagram that I would throw out in a whiteboard and we would sketch on this and kind of like a standard engineering diagram of how you draw all the different pieces of your workflow. So, you know, for example, this is pretty much a good representation of what Windsor ended up being. It was a workflow of different steps that you have as someone requests a video and then you send a video out with this

whole video generation pipeline in between, right? So, you would uh get a request, generate an audio file, do some ffmpeg processing. We also had this one GPU inference step uh which always made it annoying because we wanted to keep everything on Verscell and have everything serverless but when you have something like GPUs and Verscell just didn't have that for a long time um we had to go and sort of have this extra step and then orchestrate everything in between various functions right and so

you end up kind of having to handle your own workflow orchestration you have these cues in between that you wire everything up with for resiliency Uh and this is pretty much our actual stack. U we had this hand rolled internal tool to let you sort of human review and watch every single video that was generated and three different databases we had to set up. A lot of the state doesn't even matter to the user. It's just for our own internal admin tool and also wiring up your own

observability but stuff like sentry but we also sort of hand roll stuff before that. So, um, and it ends up being a lot of code to maintain. And I love this slide because I asked Chad GPT to try and just generate a lot of code to explain what was going on. And I actually wish it was this simple. It would have been great if all of Windsor was a single file, but it ends up being, you know, a lot more than that. You've got a whole bunch of API routes. Um, there isn't that diagram

that I have, there isn't one place I can just go read that, right? you end up having something that every time you onboard a new engineer, I have to go whiteboard and explain the same thing to them over and over again. Your code just isn't like readable. So, okay, I'm going to come back to this example, but I want to quickly switch to something else that we built at uh the workflow team pretty recently. This was something that Nate from the team started sketching out. I love this because this is an actual screenshot