I'll be upfront: the chat bot on this site wasn't my idea!
It was Matthias's, a good mate from badminton. After games we always end up talking about AI, what's actually useful and what's just noise. One day he showed me sayheyto.me, a site he built where you upload your CV and it gives you an AI version of yourself that people can chat to.
I had a go and thought, wow, that's really cool! Then I thought, I want one!
So I got to work (I didn't build it, AI did!). Different stack, my own notes, but honestly the idea is all his. I pretty much copied it.
What surprised me is that the coding was the easy bit (it's always just a prompt away). The hard bit is all the stuff around it. A few things I had to get right:
- What it knows. It only answers from notes I've written and checked myself. If it's not in there, it should say "I don't know" instead of making something up.
- How it finds things. My notes are split into small chunks so it pulls the right bit for each question, not just whatever sounds close.
- Who it is. It talks about me, not as me. It should sound like me, friendly and straight to the point, not like a corporate brochure.
- How it answers. Short answers first, real examples over buzzwords, and an honest "not my area" when it doesn't know.
- What it won't say. Some things stay private: no colleague names, no client names, nothing about other people that isn't mine to share.
- Keeping it on track. People will try to make it ignore its rules or chat about random stuff. It politely steers back to me and my work.
- Testing it. I keep a set of tricky questions and run them after every change, so a fix in one place doesn't break something else.
- Learning from real chats. I read what people actually ask and fill the gaps in my notes.
- Cost. It has limits so a bored visitor can't run up a big bill!
That last 10% took way longer than the first 90%.
None of this was new to me, which is half the fun. In September I gave a talk at a Data & AI conference here in Melbourne called "The Rise of AI Agents: From Experimentation to Operational Reality". The short version: agents are only as good as the data underneath them. Get your data foundations right first. Build the parts that make you different (your data, your rules, your tests) and just use the stuff that keeps getting better and cheaper (the models and frameworks). Start with internal tools, keep each agent's job small and clear, and only then scale up. This little chat bot is me taking my own advice.
Funny enough, getting AI to answer like a real person is exactly what Matthias does. He's been in customer support and customer success for over 15 years, and lately he's been bringing AI into support without losing the human side. Check out his website and LinkedIn.
If you like chatting with mine, go try the original at sayheyto.me!
Thanks Matthias, for the idea and the inspiration. Next round's on me!