AI Engineering.

I'm not an AI researcher, and I won't pretend to be. What I do is treat AI as a tool: I use LLM assistants to move faster on frontend work, and I've built a couple of small but real things — an agent workflow inside a mock e-commerce store, and a chatbot that lives on this site from time to time. Both are functional, neither is vaporware.

One of my favorite uses of AI has nothing to do with shipping — it's studying. I can describe a concept I want to understand, ask for analogies, worked examples, or even visualizations, and get a personalized explanation faster than any textbook. That matters a lot to me because I want to understand the data engineering pipeline inside and out, which means ML and AI are things I need to actually know, not just gesture at.

A recent example: I was listening to AI Engineering: Building Applications with Foundation Models and came across sequence-to-sequence (S2S) models — the architecture Google used to dramatically improve their translation systems, and the direct predecessor to the transformer models modern LLMs run on. S2S models use recurrent neural networks (RNNs) as encoders and decoders, and I found the concept genuinely interesting. So I asked Claude Opus to write me a practical guide to building one from scratch, tailored to a use case I was already working on (see eFrog on my GitHub).

It did. Now I'm working through it — writing RNNs and attention mechanisms by hand in Python, step by step. The immediate goal is a learning exercise; the longer-term possibility is a model that improves accuracy in one of my apps. But the broader point is that I can hand AI a subject, get a structured curriculum with examples and context built around my actual work, and learn it quickly. That's the workflow I care about.