The stack
under the vibes.
You ship software with AI agents every day. Underlay teaches the concepts underneath — Python, HTTP, servers, deployment, LLMs, agents — so you read what the AI writes, verify it, fix it, and ship like an engineer. Without becoming one.
no prerequisites · real tools · real checks · your servers
11
modules
3
tracks
30+
real checks
1
shipped capstone
the hook
Ask any LLM how many r's are in strawberry. It will lie to you.
Not because it's stupid. Because "strawberry" tokenizes as str | aw | berry— the double r hides inside a single token, and the model predicts tokens, not letters. By module 7 you'll know exactly why this happens, why models hallucinate, and how to build around it.
Start at the strawberry moduletiktoken · o200k_base · live run
3 tokens. The rrlives inside one token — the model never sees the double r. It's not hallucinating. It's token-blind.
"Hello world" = 2 tokens · english ≈ 0.75 words/token
why underlay
You will never leave without learning something.
Concepts, not bootcamp
You won't write 100% of the code. You'll understand 100% of the stack — and review what the AI produces like someone who owns it.
Real everything
Real tools, real configs, real commands, real prices. Every number on this site was verified against a live system, not a vibe.
Your servers are the lab
Every module ends with a check on infrastructure you own: a curl that returns 200, a backup restored, a model served.
The AI teaches you
Every module ships a tutor prompt. The thing you're learning to master becomes your teacher.
course 001
Vibe Coder Engineering
A concepts-first course for people who ship software with AI agents but want to understand what's underneath. Eleven modules — Python, JS/TS, Rust, HTTP & APIs, backend, deployment, how LLMs work, running LLMs, agents, and a shipped capstone — built on real tools, real configs, real commands, and real numbers. No prerequisites. The AI is both the subject and the tutor.
three ways in
track 01
1-Day Tour
~9 hours · one Saturday
A live URL: a Dockerized FastAPI app calling a real LLM, behind nginx + SSL, with git history.
H1 · Foundations
H2 · Deployment
H3 · HTTP
+5 more steps
track 02
1-Week Bootcamp
7 days · 5–6h/day
Working competence in the core stack: backend, deployment, local LLM, and a deployed RAG chatbot with an eval.
D1 · Foundations + HTTP
D2 · Python + API testing
D3 · Backend
+4 more steps
track 03
1-Month · The Real Deal
4 weeks · ~15h/week
The full 10-module curriculum: Python, TS, Rust, backend, deployment, LLM internals, local + API inference, agents — and a capstone shipped to production.
W1 · Foundations + Languages
W2 · Web + Backend
W3 · Deployment
+1 more steps
the loop
Learn. See. Run. Stamp.
Every module ends with one real check. Pass it, get the stamp. Stamps build a streak, and streaks build engineers. No endless tutorials, no certificates that mean nothing — just a chain you don't want to break.
01
Learn the concept
Plain language, one idea at a time, with the analogy that makes it stick.
02
See real code
Line-by-line reads of the actual tools the industry uses. You read, you don't transcribe.
03
Run the check
One verifiable action on your own infrastructure. curl returns 200. Migration passes. Backup restores.
04
Get the stamp
Pass the check, earn the stamp. Stamps build the streak. The streak builds the engineer.
Your first stamp is free.
Read this page, click the button. That's check #1 of the course — the streak starts now.
faq
Before you ask.
No. This is a concepts course for people who already ship with AI. You won't write 100% of the code yourself — you'll understand 100% of the stack, read real code fluently, run real checks, and review what the AI produces instead of trusting it.
None. Module 0 starts at 'what is a computer program'. Every technical term is defined the first time it appears, and a glossary lives on the site.
The curriculum is the hard part. A random conversation with an AI meanders; this course is a deliberate sequence — fundamentals → web → deployment → LLMs → agents — where every module compounds the last. The AI is your tutor inside a structure that actually converges.
The real 2026 stack: Python + FastAPI, TypeScript + Express, Rust, Docker, nginx, Let's Encrypt, GitHub Actions, PostgreSQL, Redis, Ollama/llama.cpp/vLLM, and OpenAI-compatible LLM APIs. You learn on your own VPS, not a sandbox.
Every module ends with one verifiable action — a curl returning 200, a migration passing CI, a backup restored, an eval score improving. Pass it, get the stamp. Stamps are your progress; the streak keeps you honest.
The 1-Day Tour is one Saturday. The 1-Week Bootcamp is 7 days at 5–6 hours. The 1-Month track is the full 10-module curriculum at ~15 hours a week, ending in a shipped capstone.
Pick your entry point. Leave different.
One Saturday to see the whole stack. Seven days to working competence. Four weeks to a shipped product that calls a real LLM.