Learning Roadmap

The bedrock. Skip it if you are already comfortable with NumPy broadcasting and can derive a partial derivative by hand.

• ▸NumPy vectorised ops & broadcasting
• ▸Linear algebra: matrices, eigenvectors, SVD
• ▸Calculus: partial derivatives, chain rule
• ▸Probability & statistics: expectation, Bayes, MLE
• ▸Python typing, dataclasses, protocolsUsed throughout all projects

Build the mental model for tensors, autograd, and training loops before adding model complexity.

• ▸Tensor ops, device management (CPU/GPU/MPS)
• ▸Autograd & computational graph
• ▸nn.Module, optimisers, loss functions
• ▸DataLoader, Dataset, transforms
• ▸Training loop patterns: clipping, scheduling, validation
• ▸Project: CNN image classifierFirst portfolio project

The architecture behind every frontier model. Implement each component from scratch so nothing remains a black box.

• ▸Scaled dot-product attention
• ▸Positional encoding & multi-head attention
• ▸Layer norm, residuals, feed-forward blocks
• ▸Encoder / decoder stacks
• ▸Pre-training vs fine-tuning: BERT vs GPT
• ▸Hugging Face: Trainer, tokenizers, model hub
• ▸Project: Build mini-GPTGenerative LM from scratch
• ▸Project: BERT fine-tune for NLPClassification pipeline

Sequential decision making from first principles. The math maps directly to every algorithm you implement.

• ▸MDPs, Bellman equations, value functions
• ▸Dynamic programming: policy & value iteration
• ▸Monte Carlo & temporal difference methods
• ▸Q-learning & DQN
• ▸Policy gradient theorem & REINFORCE
• ▸Actor-Critic and PPO
• ▸RL libraries: Tianshou pipeline
• ▸Project: DQN for Atari PongCNN + replay buffer
• ▸Project: PPO MuJoCo HalfCheetahCapstone

Close the gap between working code and code you can run overnight at scale.

• ▸Mixed precision training
• ▸Gradient checkpointing & memory profiling
• ▸Distributed training basics
• ▸Model export: TorchScript, ONNX, torch.compile
• ▸FastAPI + Docker inference server
• ▸Hyperparameter sweeps with W&B or Optuna