It's been a minute! The past two months have been a whirlwind of traveling and some work stuff I've been absorbed into. I haven't had a lot of time for much of anything else. A benefit to this is I've been able to step back and take an inventory of how I use my free time. Time being a limited resource I wanted to make sure I was putting it to optimal use. To justify something spending time on side projects the project either needs to provide enjoyment, ability to learn something, or some kind of return on investment. With that new lens I made a couple decisions...
So long RecipeSnap 👋
I decided to shutdown RecipeSnap. This project was no longer providing enjoyment or any educational value. Nowadays modern AI systems (ChatGPT, Claude etc.) support multimodal conversations to perform the same functions as this project. Personally I was using Claude more than my own app, so I just don't see the value. With that said there was still a ton of value here. The project started to learn more about HuggingFace and fine-tuning models. When ChatGPT launched I converted the app to leverage those APIs, but as those models have improved I've felt less of a need for this project. So, I'm thankful for the learning experience, but it's time to move on.
Building the best NBA Fan Experience on Bluesky... and more!
One area I've been investing more time into is the NBA Now feed on Bluesky. What started as a curiosity has built a dedicated audience that I want to continue to grow and expand. First, I want to copy what I've done for the NBA in other sports (F1, WNBA, MLB etc.) and ultimately expand to other topics as well. Second, I want to make the experience better for users by adding more personalization (think a "For You" feed by sport). To accomplish both of these things it was immediately clear that I needed to completely rethink the architecture.
The initial feed was a single machine the was responsible for reading the firehose, identifying if a post qualifies for the feed, and then hosting the API to serve the feed. This has worked well for the past few months when the focus was only on the NBA, but this approach isn't scalable to the other use cases.
During the past couple months I've been implementing a microservice redesign of the system that will be scalable as more topics are supported. I've been going through each part of the custom feed process and mapping processes into the appropriate service. Throughout this process I have a couple of objectives:
- Be resilient to failures so the whole feed doesn't go down. Especially if there is a corrupted record in the firehose.
- Keep up with the rate of the firehose. No falling behind! One way to do this is separating reading the firehose and processing posts into separate services communicating via RabbitMQ. If processing falls behind the processing service can be scaled up independently to keep up with the queue.
- Reduce the amount of feed-specific code to increase maintainability. Identify core concepts for filtering content for a feed and leverage those concepts across feeds instead of having code specific to each feed.
- Make it easier to support new feeds. Once implemented a new feed should be supported with just a configuration change defining the feed specific rules.
- Future proof the system. In addition to supporting multiple feeds I want to make sure I'm designing the system to support other features like better personalization and moderation in the future.
As of today the core functionility has been implemented with the new microservices and has been through some initial testing. Now I'm plotting out strategies to improve personalization and moderation to ensure I'm capturing the data I need to support those in the future. Additionally, I'm looking into how I can replace LLMs with lighter weight models to reduce costs. Ultimately, this could shift into a Phase 2 project, but it's good to have a think on it now.
Overall, this redesign has been a rich learning experience. Investing time into this is the right move to support the future growth of the project. The best part is I've been disciplined treating this as a side project and not something that is in a constant state of needing to be done immediately. I'll have more to share in the future on this!
What I've been Reading 📚
Foundation and Earth- Finally finished this months later. The ending did just enough to make me want to read the next in the series. But first...
Wind and Truth Finally getting started on this! I'm slightly regretting getting the hardcover version of this because this thing is a brick! It's been on my desk for months and now is the time to finally get absorbed back into the Cosmere universe!
The Charisma Myth- More investment in soft skills.
**AI Engineering- There is so much information in this book. Must read if you're working with LLMs.