Projects

8 capstones · one per unit · Google Colab

• Foundations

  Spam Classifier

  ~45 min

  Train your first text classifier on real SMS messages

  What you’ll make

  A logistic regression text classifier, compared with a Naive Bayes baseline, trained on the UCI SMS Spam Collection. You will turn messages into TF-IDF features and evaluate predictions with confusion matrices, precision, and recall.

  • Logistic Regression
  • TF-IDF
  • Precision/Recall
  • Confusion Matrix

  Before you startBasic Python. A little curiosity about data.Machine Learning FoundationsView project →Notebook ↓ for Spam Classifier
• Classical ML Algorithms

  Customer Churn Predictor

  ~120 min

  Evaluate a churn model and explain its limits

  What you’ll make

  A runnable notebook, saved preprocessing/model pipeline, environment instructions, dataset and split manifest, error analysis and learner-authored decision memo. Compare a baseline and three candidates on validation F1, freeze the choice, report final-test results, and inspect SHAP associations without claiming causal effects.

  • XGBoost
  • Random Forest
  • SHAP Explainability
  • F1 Score
  • Class Imbalance

  Before you startBasic Python. A little curiosity about data.Machine Learning FoundationsView project →Notebook ↓ for Customer Churn Predictor
• Architectures

  CNN Visual Inspector

  ~50 min

  Fine-tune a ResNet and see what it actually looks at

  What you’ll make

  A fine-tuned ResNet50 trained on 10 food categories from Food-101. You'll apply Grad-CAM to generate visual explanations for any prediction, and test your model on a photo of your choice to see what the network is actually "looking at."

  • Transfer Learning
  • ResNet50
  • Grad-CAM
  • Data Augmentation
  • Fine-Tuning

  Before you startMachine Learning Foundations and basic Python.Deep LearningView project →Notebook ↓ for CNN Visual Inspector
• RNNs & Sequence Models

  LSTM Creative Writer

  ~45 min

  Train a network to write like Sherlock Holmes

  What you’ll make

  A 2-layer stacked LSTM trained character-by-character on The Adventures of Sherlock Holmes (Project Gutenberg). You'll generate text at three different temperatures to control creativity vs coherence, and observe how the vanishing gradient problem manifests as the model's "memory" degrades over long sequences.

  • LSTM
  • Character-level Modelling
  • Temperature Sampling
  • Vanishing Gradients
  • PyTorch

  Before you startMachine Learning Foundations and basic Python.Deep LearningView project →Notebook ↓ for LSTM Creative Writer
• Transformers & LLMs

  Financial Sentiment Fine-Tuner

  ~55 min

  Fine-tune DistilBERT on the data hedge funds use

  What you’ll make

  A fine-tuned DistilBERT model (66M parameters) trained on 4,845 labelled financial sentences. You'll visualise the confusion matrix, inspect where the model gets confused (neutral vs negative is the hard case), and optionally publish your model to Hugging Face Hub where it gets a permanent URL you can link from your resume.

  • DistilBERT
  • Hugging Face Transformers
  • Fine-Tuning
  • Financial NLP
  • Confusion Matrix

  Before you startComfortable with Python. ML fundamentals recommended.LLM FoundationsView project →Notebook ↓ for Financial Sentiment Fine-Tuner
• Generative & Frontier

  VAE Latent Space Explorer

  ~55 min

  Build the latent space that powers image generation

  What you’ll make

  A Variational Autoencoder trained on MNIST with a 2-dimensional latent space. You'll produce: (1) a reconstruction grid, (2) a 2D scatter plot of the latent space coloured by digit class, (3) smooth interpolation between two chosen digits over 12 steps, and (4) a 10×10 generative grid sampled from a uniform latent grid. You'll also run a KL ablation to see what the regularisation term actually buys.

  • Variational Autoencoder
  • KL Divergence
  • Reparameterisation Trick
  • Latent Space
  • PyTorch

  Before you startMachine Learning Foundations and basic Python.Deep LearningView project →Notebook ↓ for VAE Latent Space Explorer
• MLOps

  Local ML Serving & Monitoring

  ~150 min

  Persist a model, test an API, investigate an incident

  What you’ll make

  A reproducible notebook and local project with MLflow run artifacts, a FastAPI app, executable API tests, environment instructions, a decision memo and an incident runbook. Measure sequential in-process request latency and inspect drift. Registry promotion, external deployment and concurrent load testing are separate later work.

  • MLflow
  • Validation Selection
  • FastAPI
  • API Failure Tests
  • Data Drift Detection
  • Local Latency Measurement

  Before you startExperience training and evaluating a model.MLOps & ML EngineeringView project →Notebook ↓ for Local ML Serving & Monitoring
• Responsible ML

  Fairness Audit

  ~50 min

  Reproduce a real biased algorithm — then fix it

  What you’ll make

  A complete fairness audit of the COMPAS dataset using the Fairlearn library. You'll compute FPR, FNR, and calibration metrics split by race, visualise the disparity, apply Exponentiated Gradient (in-processing) and Threshold Optimizer (post-processing) mitigations, measure the accuracy-fairness trade-off, and write a verdict as if submitting to a regulator.

  • Fairlearn
  • False Positive Rate
  • Equalized Odds
  • Threshold Optimization
  • Disparate Impact

  Before you startBasic Python. A little curiosity about data.Machine Learning FoundationsView project →Notebook ↓ for Fairness Audit