Day 8: StarLens Product Hunt launch, Tackling LLM Rate-Limits [Update #04]

Max here and it's the end of day eight of the AI Sprint today the big headline is that I launched styland on product hunt it's been quite a ride today in the morning I started monitoring it and I started to see some errors coming in thanks to the error workflows I had set up I was getting slack notifications about these and so the first part of the day was just triaging these and fixing them there was a few different buckets of issues people were submitting things like the entire profile and not just the username the cool thing about that was with Ned I was able to fix erors like that pretty quickly with an if node checking for that condition with a little conditional rejects I don't have as many in-depth updates for today cuz I was kind of heads down fixing issues as they came up so that was one small bucket of errors but the biggest bucket of Errors I was dealing with was actually rate limits from the different large language models I was using because they're all pretty heavily rate limited these days the eror work was helped me there I was able to identify the pattern pretty quickly that we've got this rate limit issue and so one of the first things I did was change the default model from Claude to llama that helped on the rate limiting side with Claude and then I was getting some rate limits from llama and since I'm using grock there was a few different llama models that were pretty similar I was able to use click a drop down change a model and hit save really helped me uh reduce how many errors I was getting but it was really humbling because there was some spikes throughout the day we had sometimes 60 Folks at a time trying to request the cool thing is that the n and end workflows itself uh handled that pretty find it was just failing due to the rate limit errors I was talking with our CEO Yan and we were discussing how you know if this was a production environment of course we could have some sort of round robin cycling on those

credentials and I think that's such a great example of how flexible idend can be a few moments later stalins is live we're getting quite some volume of traffic so we're starting to getting a few errors in I'm triaging those one of the ones that's coming in I was looking at this error execution here and I noticed that we've got an error in the respon to we web node so what I did is Click debug and editor this moved all the data into my flow here and so I can inspect it I can click on respond to we web and I can see if we go down in here that I'm doing a two Json string and projects to follow is null so what we're going to do here is go to chat GPT and current so let's do that let's go back in the workflow let's do that um and then here we're going to set null instead so now we've fixed uh this specific eror and I can continue monitoring other errors that come in um and basically triage those cases so let's have a look at this one here believe this is the same case yep so we just fixed this case so we should not see this case again yay 12 seconds later so for the rest of the day I'm going to be monitoring stallins as the west coast and the US wakes up and then as I sign off for tonight and

get into Day N tomorrow we're going to pick up a new project and one of the things I'd love to do I think rag is a big topic but most focus on Vector stores now it's retrieval augmented generation so you're retrieving something to basically ground the AI in some sort of truth that doesn't have to only happen with the vector store and I think tools are a great example so in some of the workflows I was building for sty lens the AI agent was leveraging tools where it could for example search GitHub repos we want to make an example of that taking some applic notion and showing how you can create a notion tool which it can search database Pages without having to put them in a vector store I think that's going to reduce the barrier to entry for a lot of folks I'll catch you tomorrow till the next AI Sprint day