Engineering
The path · foundation to applied
Foundations first: what a shell actually is, how processes and file descriptors work, and the expansion rules that explain nearly every bash surprise.
Everything that happens before your app serves a single byte — what each piece of the machinery is for, then the commands that drive it.
The tool you touch every day, and the one with no guide yet — branching, merging, history, and the object model underneath.
Not the basics — the intermediate-to-advanced language, since you already write both. Python's data model, decorators, generators, async, and typing; TypeScript's type system in earnest — generics, conditional and mapped types, inference and narrowing. Idiom, and the sharp edges.
The browser decides where every box goes by rules you were never shown — learn them from the box out, until you can predict any layout and build a nav or a menu without fighting it.
The browser runs your code, measures the page and repaints on a loop you don't control — learn that loop, its rulers and its pointers, until you can build a pannable, zoomable canvas the way React Flow does.
Everything between a browser and your database, taught as one machine — the socket, the event loop, the proxy, the pool, the query planner and the cache, and why each layer had to exist.
How a request actually crosses the internet — only partly covered inside the backend and Linux guides today.
The habit that separates "works on my machine" from strong engineering: unit, integration, and end-to-end, and what is worth testing at all.
Sessions, tokens, hashing, and the common vulnerabilities. The day you ship to real users, you are holding their data.
Every core data structure built and priced from scratch in Python — and the decision table that tells you which one a problem actually wants.
Every core algorithm built and measured from scratch in Python — recursion, sorting, graphs, dynamic programming — and how to read a new problem for the pattern it wants.
Every layer of a real system — storage, replication, consistency, caching, failure — measured from scratch against a live Postgres and Redis, and how to read a design problem for the force it turns on.
Optional, for later: how code runs on the CPU and in memory, and concurrency — threads, async, and the races between them.
Mathematics
Parallel · strengthens the AI work
Counting, fractions, powers and the whole of school algebra — every rule earned before it's stated, with a picture for each one.
Geometry, functions, growth, trigonometry, chance and the edge of calculus — the shape behind each idea, drawn before the formula.
The highest-leverage maths for AI: embeddings and attention are vectors and matrices. Nothing in the two maths guides touches it — the one real maths gap for where you're headed.
Shape and change reaches the edge — rates and areas; this is the subject proper: limits, derivatives, integrals, the fundamental theorem. Mostly matters if you go down to building models.
Optional: sets, logic, proof, combinatorics, graphs — the maths computer science itself rests on.
Engineering is the path · mathematics strengthens it
Applied AI
Where the two lanes meet, and the destination. Read as a ladder: you build products at Level 1, sharpen them with the intuition of Level 2, and only descend to Level 3 if a model itself becomes your product.
Level 1 — AI engineer. Where products get built, and where your edge lives.
Build the agent loop by hand, then with a framework, then wire it into a real web app — seeing exactly what each layer of abstraction does for you, and when a protocol like MCP earns its place.
Whether a change to your agent actually helped is a statistical question — traces, assertions, the sample sizes that make a pass rate mean something, and judging with a model you have validated.
Give a model the right context: vector search, chunking, grounding. The practical core of most AI products, beyond what the agents guide covers.
Cost, latency, streaming, guardrails, evaluation at scale — turning an agent that works into a product people rely on.
Level 2 — mechanistic intuition. The mental model that makes frontier work possible. The ideas, not the training maths.
Tokens, embeddings, attention, sampling, and why they fail. The mental model that separates frontier work from cargo-culted prompts. The ideas, not the training maths.
Level 3 — model builder. Optional, later — only if a model becomes your product.
Optional, later — only if you want to fine-tune, or a model itself becomes your product. How models are trained, not just used.