Learn ML

A 12-episode video series that teaches ML engineering by building real projects.

This isn't "here's how to write a for loop." This is "here's what you need to understand to build ML systems" -- the concepts, the intuition, and the hands-on experience that AI coding tools can't replace.

The Format

Every episode has two parts:

The Knowledge (5-10 min) -- Concept explanation, intuition, and mental models. Interactive slides you can follow along with in your browser.

The Work (10-15 min) -- Build something real. Step-by-step Python scripts where each one demonstrates a concept with clear input and output.

Episodes

#	Topic	Project	Slides	Video
1	PyTorch	Pokemon card price predictor	Slides	Coming soon
2	Transformers	Build a mini-GPT	Slides	Coming soon
3	Hugging Face	Model comparison tool
4	Embeddings	Searchable speech database
5	Fine-tuning (SFT)	Domain-specific assistant
6	LoRA / QLoRA	7B model on consumer GPU
7	DPO	Align your fine-tuned model
8	Evaluation	Domain evaluation suite
9	GPU Optimization	2x training throughput
10	Distributed Training	Scale beyond one GPU
11	Reinforcement Learning	Game AI agent
12	Triton & Kernels	Understand ML infrastructure

Series intro slides

Getting Started

Requirements

Python 3.11 or 3.12
An NVIDIA GPU with CUDA support (tested on RTX 3060)
~2GB disk space for PyTorch

Setup

Create a virtual environment and install dependencies:

python -m venv .venv

# Windows
.venv\Scripts\activate

# Linux / Mac
source .venv/bin/activate

# Install PyTorch (with CUDA)
pip install torch --index-url https://download.pytorch.org/whl/cu124

# Install other dependencies
pip install pandas matplotlib scikit-learn pyarrow requests

Running an Episode

Each episode lives in its own directory under episodes/. Scripts are numbered and meant to be run in order -- each one builds on the previous.

cd episodes/01-price-predictor

# Step 1 produces output that Step 2 needs, and so on
python step1_tensors.py
python step2_model.py
python step3_training_step.py
python step4_train.py
python step5_evaluate.py

Viewing Slides

Open any .html file in the slides/ directory in your browser. Navigate with arrow keys, spacebar, or the buttons in the corner. Touch/swipe works on mobile.

Episode Details

Episode 1: PyTorch

Concepts: Tensors (shape, dtype, device), autograd, the training loop, nn.Module, DataLoader

Project: Train a neural network to predict Pokemon card market prices from card attributes (rarity, type, HP, etc.) using real data from the Pokemon TCG API.

Script	What it does	Output
`fetch_data.py`	Pulls 2,000+ real Pokemon cards from the API	`pokemon_cards.csv`
`step1_tensors.py`	Convert CSV to tensors, inspect shape/dtype/device	`card_tensors.pt`
`step2_model.py`	Build the neural network, explore parameters	--
`step3_training_step.py`	One training step: forward, loss, backward, update	--
`step4_train.py`	Full training loop, 100 epochs	`trained_model.pt`
`step5_evaluate.py`	Test predictions, metrics, plots	`results.png`

Episode 2: Transformers

Concepts: Attention (Q/K/V), multi-head attention, causal masking, positional encoding, the transformer block (feed-forward, residuals, layer norm), decoder-only generation

Project: Build a mini-GPT from scratch and train it to generate text. We trained ours on One Piece episode synopses.

Note: You need to provide your own training text. Place it in episodes/02-mini-gpt/input.txt. Any plain text works -- novels, articles, wiki pages. Aim for 500KB-3MB.

Script	What it does	Output
`step1_attention.py`	Implement attention from scratch, visualize weights	`vocab.pt`, `input.txt`
`step2_multihead_mask.py`	Multi-head attention + causal mask	`config.pt`
`step3_transformer_block.py`	Build the full model: blocks + embeddings + output	`model_config.pt`
`step4_train.py`	Train on your text (~5-15 min on GPU)	`trained_minigpt.pt`
`step5_generate.py`	Generate text from prompts, temperature comparison	`training_loss.png`

Philosophy

Concepts over code -- Understand what's happening. AI writes the boilerplate.
Real projects -- Not toy examples. Actual useful things.
Step by step -- Each script demonstrates one concept with clear input and output.
AI-resistant skills -- Systems understanding, debugging intuition, domain expertise.

License

MIT