Sidekick AI

Over the past months, I've put hundreds of hours into redesigning the website and building a backend from scratch using Rust, the Rocket framework, and Sqlite. I'm happy to say that today I finally finished the last feature I wanted before launch (the website embed).

I've been building the systems to scale and grow in the future, so it definitely took longer than just hacking them together so it works, but I think this should scale well and be able to be picked up by other devs and expanded upon in the future, so I'm feeling pretty good about the state of the code.

Here's to hoping my next post will be about my first customers!

I finally decided to choose a name, design a theme, and launch a website. I've done a little web dev before, so I decided to build the website from scratch, using a Linode server and the flask web framework. Since then, the website has undergone changes almost every single day, but this first crappy website was an important beachhead onto the internet. Around this time, I submitted for Sidekick AI LLC to be officially incorporated.

Working on various ML projects, I started figuring out what modern ML was good for and what it wasn't. I felt that chatbots were a promising area, and that they could eventually improve to the point of being indistinguishable from humans. To me, this seemed too big of an opportunity to pass up, so I started working on a line of chatbots that improved with each iteration. Eventually I got to a model that could hold a conversation (a verrrrry short and basic one).

While I was working, I was searching for what my bots can do, how I could make this a business. I thought back to every time I had to scour a company's website trying to find an email for their customer service, only to realize none exists. I remembered the hours I spent on hold waiting for a depressed call center worker to answer my call and promptly put me on hold again. This seemed to be an industry that had no ability to scale to modern needs, because it was entirely reliant on human labor, which is expensive. Not only could it not scale, it had a high bar of entry. Hiring a full time customer service agent is expensive. The alternative is making customers wait hours or days for someone to get back to them.

This suddenly seemed like the most obvious and influential application of ML chatbots, so long as they can give customers a better experience than the hardcoded ones, which are less than useless. Quickly I came up with the business model: chatbots as a service, accessed through a variety of channels. I got to work building more customer service oriented chatbots.

Back in 2019, I started working on what would become the Sidekick AI codebase. I didn't know exactly what I was working on, but I was learning loads about machine learning and natural language processing.

I started with basic image classification, and gradually moved up to language models, Seq2Seq models, and then Transformer models. Along the way, I made plenty of projects, 90% of which were boilerplate, so I started writing some of this boilerplate into a library, which eventually morphed into the first iteration of the Sidekick AI experimental library!