Projects you can run tonight
Each project is small enough to finish in an evening, and complete enough to actually run: every script executes end to end on a CPU or Apple Silicon in under 15 minutes, with a checklist of what to look for when it works.
An MNIST-class CNN in PyTorch on CPU
Train a two-layer convolutional network on 16x16 digit images in under a second on a laptop CPU, read the loss curve, print a confusion matrix, and find out which digits it gets wrong.
A tabular classifier with a FastAPI endpoint
Train a scikit-learn pipeline on the breast cancer dataset, save it with joblib, serve it from a FastAPI app with /predict and /health, and test the app in-process before you ever start a server.
Fine-tune a small model with LoRA on Apple Silicon in under 15 minutes
Teach SmolLM2-135M a date-format conversion it cannot do out of the box, with a rank-8 LoRA adapter trained for 60 steps on MPS or CPU, then save, reload and verify the adapter.
A tiny RAG system over a folder of text with local embeddings
Chunk a folder of notes, embed the chunks with all-MiniLM-L6-v2 on your CPU, retrieve with a cosine index in plain numpy, then hand the top chunks to a local model through Ollama for a cited answer.