Projects
Every concept is reinforced by building something real. Projects include starter code, grading rubrics, and stretch goals.
CNN Image Classifier
Build a ResNet-18-style CNN from scratch, train on CIFAR-10, and target >90% accuracy.
Mini-GPT (Shakespeare)
Character-level GPT with causal attention and sampling. Trains on a single GPU in about 20 minutes.
BERT Sentiment Fine-tune
Fine-tune BERT on SST-2 using Hugging Face Trainer, with an optional LoRA branch included.
DQN Atari Pong
Full DQN with replay buffer and target network. Train an agent to beat the Pong opponent.
PPO MuJoCo HalfCheetah
Implement PPO + GAE and train a HalfCheetah-v4 agent with logging and video recording.
Tianshou CartPole Pipeline
Use Tianshou to build a reusable RL training pipeline with collectors, replay buffers, and trainers.
FastAPI Inference Server
Wrap a trained PyTorch model in FastAPI, add batching, Docker packaging, and a health endpoint.