12 Episodes — Series Overview

ML Engineering
Video Series

Learn to build ML systems — not by memorizing code, but by understanding the concepts well enough to direct AI tools to do it for you.

The Angle

This is not a coding tutorial

AI tools like Claude, Codex, and Cursor can write the boilerplate. What they can't do is decide what to build, debug why it's broken, or know which approach fits your problem.

Not this

"Here's how to write a for loop in Python"

This

"Here's what you need to understand to direct Claude to build ML systems"

We focus on the skills AI can't replace: systems understanding, debugging intuition, and domain expertise.

The Format

Every episode, two parts

Part One

The Knowledge
5–10 minutes

Concept explanation and intuition. What is this thing, why does it exist, when would you use it, and what are the mental models that make it click.

No code in this section — just understanding.

+

Part Two

The Work
10–15 minutes

Build something real. Not a toy example — an actual useful project that applies the concepts. You follow along and see the process, not just the finished result.

This is where it becomes tangible.

Target length: 15–25 minutes per episode

The Format

The Knowledge — what it covers

Each Knowledge section is built to give you a durable mental model — something that stays useful even as tools and libraries change.

  • 01

    What it is — the concept in plain language, no jargon walls

  • 02

    Why it exists — the problem it solves, and what came before it

  • 03

    How it works — the core mechanism, with intuition over formulas

  • 04

    When to use it — practical decision-making, tradeoffs, and gotchas

What you won't see

  • Walls of math notation
  • Line-by-line code walkthroughs
  • Theory disconnected from practice

What you will see

  • Analogies that make concepts stick
  • Visuals and diagrams
  • Real-world context for every idea
  • The "why should I care" for each concept
The Format

The Work — what you'll build

Every episode has a hands-on project that applies the concepts to a real problem — not contrived exercises, but tools and systems you'd actually want to use.

Projects are grounded in real domains:

Project Domains

  • Prediction markets — Kalshi, Trump speeches, sentiment
  • Crypto — Solana, Pokemon card perp dex, market data
  • Game AI — 4X NFT strategy game, reinforcement learning

How Work sections are structured

  • Step-by-step — build incrementally, not all at once
  • Show the process — including mistakes, debugging, and decisions
  • Use AI tools — demonstrate working with Claude/Copilot naturally
  • End with something real — a working model, tool, or system

Projects build on each other — data from Episode 5 gets reused in 6, the model from 5–6 gets improved in 7. By the end, you've built a full ML pipeline.

The Roadmap

12 episodes, one arc

01
PyTorch
Build a price predictor from scratch
02
Transformers
Build a mini-GPT
03
Hugging Face
Model comparison tool
04
Embeddings
Searchable speech database
05
Fine-tuning
Domain-specific assistant
06
LoRA / QLoRA
7B model on consumer GPU
07
DPO
Align your fine-tuned model
08
Evaluation
Domain evaluation suite
09
GPU Optimization
2x training throughput
10
Distributed Training
Scale beyond one GPU
11
Reinforcement Learning
Game AI agent
12
Triton & Kernels
Understand ML infrastructure
Foundations
Using Models
Customizing
Measuring
Scaling
Advanced
Philosophy

Four principles

01
Concepts over code

Understand what's happening under the hood. AI writes the boilerplate — you need to know what to ask for and whether the output makes sense.

02
Real projects

Not toy examples. Every episode builds something you'd actually use — a price predictor, a search engine, a game AI. Grounded in crypto, trading, and gaming domains.

03
Tutorial style

You follow along and see the full process — including the debugging, the decisions, and the dead ends. Not line-by-line dictation, but a realistic workflow.

04
AI-resistant skills

We focus on what AI tools can't do for you: systems thinking, debugging intuition, architecture decisions, and domain expertise. The stuff that makes you an engineer, not a typist.

Audience

Who this is for

  • Developers who want to become ML engineers — you can code, now you want to understand the ML side

  • Builders using AI coding tools — Claude, Codex, Cursor — who want to give better instructions and catch mistakes

  • People who learn by building — not by watching lectures or reading textbooks

Assumes

  • You can write code (Python preferred, any language fine)
  • You don't need ML explained from zero — we start at PyTorch, not linear algebra
  • You're willing to get your hands dirty with real projects
▲ ▲ ▲ 12 episodes — let's go

Start with PyTorch

01 / 09