My AI Journey

I occasionally get asked how I learned "all this AI stuff." The honest answer is that I never sat down with a grand plan to become an AI expert. I was just an engineer who liked solving problems and following interesting questions wherever they led.

In the beginning, I wasn't interested in using AI to write emails or generate images. I wanted to know whether these models were actually thinking, or whether they were simply very sophisticated autocomplete engines. That led me down a rabbit hole of experiments. I'd ask deliberately tricky questions, explore edge cases, give them logic puzzles, math problems, and scenarios designed to expose weaknesses. Sometimes they surprised me. Sometimes they failed in spectacularly interesting ways. Both outcomes taught me something.

As my curiosity grew, I realized I couldn't keep treating AI as a black box. I wanted to understand what was happening under the hood. That meant learning about transformers, tokens, embeddings, context windows, inference, and enough machine learning to understand why these systems behaved the way they did. I wasn't trying to become a machine learning researcher. I simply wanted a mental model that matched reality instead of marketing.

Eventually, reading articles stopped being enough. Like most things in software engineering, I learned far more by building than by consuming information. I wrote little experiments. I hosted local models. I generated images. I built writing assistants. I automated repetitive tasks. Every project taught me something new, and nearly every project forced me to learn a technology I hadn't expected to need.

Running models locally was another major turning point. Once you're responsible for making the whole stack work yourself, you stop thinking only about prompts. Suddenly you're learning about GPU memory, quantization, inference engines, context sizes, performance bottlenecks, CUDA, and why one model behaves completely differently from another. AI became less like a website I visited and more like infrastructure that I operated.

That shift changed the kinds of questions I asked. Instead of wondering which prompt was best, I became interested in building systems around models. I started exploring agentic workflows, tool calling, Model Context Protocol (MCP), orchestration, memory, evaluation, and observability. The model itself gradually became just another component in a much larger engineering system.

Somewhere along the way, I became fascinated by memory and knowledge. Not just how to make a chatbot remember previous conversations, but deeper questions: What should an AI remember? What should it forget? How does experience become knowledge? How do multiple agents cooperate? The more I explored those questions, the more they resembled questions about human cognition as much as software engineering.

Only after all of that did I begin applying what I'd learned to production systems. By then, I wasn't building AI because it was fashionable. I was building tools that made developers, QA engineers, product owners, and other teams more effective. The AI itself wasn't the product; it was one piece of a larger solution designed to reduce friction and help people accomplish real work.

I've had to pick up a lot of different technologies and concepts, and I'll continue to have to because there are new things popping up every day. My discovery process has stayed pretty consistent: I found something that made me curious, I learned enough to build a small experiment. That experiment taught me something I didn't know. I learned a little more, built something slightly larger, and repeated the cycle.

If someone asked me today where to begin, I probably wouldn't hand them a stack of papers or a list of YouTube videos. I'd encourage them to pick a problem they genuinely want to solve, learn just enough to build a first version, and then let curiosity do the rest.