AI assistant LLM
AI assistant - MikeGPT
This is a Large Language Model like ChatGPT that I trained on my own emails and website. It also has the capability to forward me a message. If you ask it to contact me, it will prompt you for your name and email, and forward the message to me.
Technical Details
This is a custom large language model that I fine-tuned on years of my own emails, and on data from this website, so it writes the way I do. It knows about my work and skills.
The base model is Qwen 2.5, a capable open-source model, which I trained further on my own data. It is not a generic chatbot wired up to someone else's service: I host the model myself on rented GPUs (RunPod), and a small Cloudflare Worker sits in front of it to handle security and keep costs in check.
Because it's running on a fairly high-end GPU I have it shut down when no one is using it. It may take a moment for the first response. Other than that small delay, this is a commercial-ready example of training an LLM and giving it a web interface.
What it is for
Mostly, it's just a fun demo! But you can ask it about my skills, or tell it about a project you might want to create. It will act as a secretary and forward the details to me so that "real me" can get back to you.
Keep in mind that, like any AI chat, it can malfunction. I have trained many of the most common malfunctions out of it. But, if something it says sounds odd, you can always contact me directly.
How it works
When you send a message, it goes first to a Cloudflare Worker, a small program running at the network edge. The Worker makes sure you're not a bot and applies rate limits. Then it passes your message to the model running on RunPod, which writes the reply. If you ask to be contacted, the Worker emails your details to me. To save money, after an idle period it will shut down and await the next chat.
Privacy
Your chat messages are not stored anywhere, and if you leave your name, email or phone number it is used only in an email message forwarded to me, never stored.
I pseudonymized the training data before using it, so the real names, emails, and phone numbers were not trained. The model weights stay on private infrastructure.
Give it a try.
Ask it anything about my work, or tell it about your project and let it pass your details along. If you would rather reach me the usual way, the contact page works too.