From API calls to AI engineer: what the journey actually taught me
Everyone starts by calling a model API and feeling like a wizard. The real skill curve starts after that: the math, the evals, and the engineering around the model.
The warm-up lap
My first "AI feature" was an API call with a prompt taped to it. It worked in the demo and fell apart the first time a real user typed something weird. That gap, between a cool demo and a product that survives strangers, is what AI engineering actually is.
The journey since has been layers: prompting, then retrieval, then agents and tool use, then the unglamorous parts nobody posts about: rate limits, caching, cost budgets, and guardrails against prompt injection. The chatbot on this site runs all of those lessons in about three hundred lines.
The math that actually matters
You can ship AI products without touching the math, but you hit a ceiling fast. Linear algebra is the language underneath everything: an embedding is just a vector, similarity is a dot product, a transformer layer is matrix multiplication at industrial scale. Once vectors and matrices feel physical, retrieval and fine-tuning stop being magic words.
Statistics is the other half, and honestly the more useful one day to day. A model is a probability distribution over tokens. Temperature reshapes that distribution. An eval is a sample, and if your eval set is ten questions you are reading tea leaves, not measuring anything. Understanding variance turned my "it seems better" into confidence intervals.
Linear algebra explains what the model is. Statistics explains whether you can trust it.
The skills stack today
Where the journey has landed so far: building with the Claude API and other LLM APIs, streaming and tool use, retrieval pipelines, agent design with persistent memory, prompt-injection defenses, and cost engineering with prompt caching and rate limits. Plus the classic full-stack work that wraps it all, because an AI feature without a product around it is a party trick.
The Zura coach inside ZuraLog is the biggest application: it takes fitness, sleep, nutrition, and heart data and turns it into daily guidance. Different pillar, same discipline: ground the model in real data, cap what it can claim, and measure the outputs.
Shipping beats theory
Every meaningful lesson came from production, not tutorials. Users found injection attacks my tests missed. Costs spiked where I assumed they would not. Latency mattered more than intelligence in half the features. Reading papers is useful, but deploying something and watching strangers use it teaches faster.
The debrief
If you are starting the same journey: build something real immediately, learn the linear algebra when the black box starts to bother you, and learn statistics before you trust your own evals. Everything else, the frameworks, the buzzwords, rotates every six months. The math and the shipping habit compound.