This is the first real post on the new Studio303. I figured I'd start by talking about something I built, got excited about, and quietly retired - an AI chatbot called Tripp. More than a year ago now I finally climbed out of a life-sized pit and had some energy to engage with the larger world again and started really getting excited about AI and LLMs. And while I agree there are very serious issues surrounding these products (environmental impact, copyright madness, and the effects on our collective brainpower among them) nevertheless I find AI fascinating (as a technology), exciting to use, and genuinely useful.
I wanted to get my hands dirty building something with AI - not just using it - and landed on making a RAG chatbot that could be trained on a website’s data but also on internal documents. The idea was that we could seriously limit one of the major drawbacks of AI: reliable accuracy (i.e., its tendency towards fabrication/hallucination).
I used Vercel’s AI toolkit which provides a basic chat interface and you wire into whichever LLM API you want and there you go, a chatbot. I have to say that it was a nice touch, giving us that early win in development. By building this I came to understand, at least a little better, how to manipulate queries to the LLM, about vector databases, and all about the myriad ways we might choose to chunk and parse our text. I’d have to say it was this last thing that showed me that the dream in my head of a hyper-accurate couture chatbot was nowhere near as (relatively) simple as it seemed. Despite various tweaks perfect accuracy (even on a small set of information) was fleeting.
At this point I wanted to get an MVP out there to accompany my freelance web development offerings. The idea: sell the chatbot as an add-on service, handling all the messy work of identifying the right data, figuring out the format, and ingesting god-knows-how-much of it into the vector database. The problem is, was, and probably always will be, that I am a terrible terrible salesperson. I love building with code, creating designs and really performant websites, it’s a fantastic challenge that is satisfying to see through to the end. But, with a full time job, spouse, and two kids, my energy and time for grinding cold calls or whatever is less than zero. Also, such work is not amenable to my personality at all. But anyway, I wanted a viable product to put on my site (this one, though an older version). It was almost completely accurate at this point (though of dubious utility for such a small site).
The next problem was one of packaging. My plan was to integrate Tripp directly into any Next.js site I was building for clients - not the most scalable solution, but practical enough (at my small scale). What I didn't expect was that integrating it into an existing site (the Studio303 homepage) running a different Next version with breaking changes and a completely different router would be such a headache. It was. Why didn’t I think it would be? I can’t say.
And while I moved onto other projects I left my newly redesigned site and the chatbot sitting there, gleaming in the sun (at least in my mind) but big surprise: no marketing, no selling tends to lead to no nothing!
Moving Forward
As I continued to follow AI and the marketplace it became clear to me that there were real companies—big, small, startup, whatever—trying to really perfect RAG chatbots for websites. It also became clear to me that this type of functionality was quickly becoming “solved” in the sense that there would soon be products well better than mine that would be easy to use. I could always just use those for my clients. I love building things but if something I make is supposed to be useful but just isn’t, well, then what?
And now I have pivoted this website again and I'm letting myself follow a passion of building and shipping small apps: projects, tools, games, whatever exciting idea I have and feel I can make a reality. It’s also a pivot to actually being vulnerable and sharing things on the web. It’s not something I’ve ever really done. Any social media or blog posts I’ve ever written abided by certain obscure privacy oriented or protection-seeking impulses. And so, I built an agency site for work that I didn’t really want to do and started making blog posts about topics I didn’t care about. I’m excited to try something new.
Thanks for the fun (and the insight into AI) you gave me Tripp!
(Tripp was named after Matthew McConaughey’s character from ‘Failure to Launch’ for no other reason than it popped into my head and seemed like a good idea at the time.)