AI Engineering
An ongoing track where I'm learning and building AI from the ground up — implementing the core ideas by hand and wiring them into real automation.
This is the thread I keep coming back to: actually understanding how modern AI works by building it, not just calling APIs. I'm working through it from the ground up — implementing the core ideas by hand first, then learning the production tools on top — and turning what I learn into small pieces of automation along the way.
What I'm building
The path runs bottom-up. The math foundations and classical ML at one end; transformers, LLMs, and agents at the other. I write each algorithm from raw math before reaching for a framework — backprop, a tokenizer, attention, an agent loop — so that by the time PyTorch or an agent SDK shows up, I already know what it's doing underneath. Every piece ends as something reusable: a prompt, a small tool, an implementation I can point to.
How I'm approaching it
The loop is the same each time: read the problem, derive the math, write the code, run it, keep the artifact. The "build it, then use it" split is the part I find most valuable — implementing a smaller version myself, then running the same thing through the real library, which is where it finally clicks. Alongside the learning, I lean toward automation: taking a concept I've just understood and wiring it into a script or agent that does something useful without me in the loop.
Why
Models keep getting better; what compounds is knowing what to build and how the pieces actually fit together. I'd rather be the person who can reason about a loss curve or an agent's failure mode than someone who can only wire APIs together. This is the long game I'm investing in.