Home / Projects

Build it the way industry builds it

Ten end-to-end projects across machine learning, deep learning, computer vision, NLP and LLMs. Each one is built like production work by an experienced AI engineering team, and every step is shown and explained: nothing is skipped. You get the full repository, the reasoning behind each decision, and checks to confirm your own version works.

The projects

AvailableProject 1

Will the patient show up?

Predict missed medical appointments, turn risk into a cost-based call list, and ship it as a tested API.

Machine learning · classificationBeginner15–20 h
AvailableProject 2

What's this house worth?

Estimate Melbourne house prices with an honest 80% range, leakage-reviewed features and error analysis by area.

Machine learning · regressionBeginner15–20 h
AvailableProject 3

Who are our customers?

Segment shoppers from the UCI Online Retail data, choose k by stability, and prove segments beat RFM rules on next quarter.

Machine learning · unsupervisedIntermediate15–20 h
AvailableProject 4

Tomorrow's electricity demand

Forecast tomorrow's hourly Queensland power demand from public AEMO data, with a walk-forward backtest and calibrated ranges.

Machine learning · time seriesIntermediate20 h
In developmentProject 5

A CNN from scratch

Build and train a convolutional network in PyTorch, with augmentation, loss-curve reading and debugging.

Deep learningIntermediate20 h
In developmentProject 6

Plant disease detector

Transfer learning on PlantVillage leaves, Grad-CAM explanations and a test on real-world photos.

Computer visionIntermediate20 h
In developmentProject 7

Spot the safety gear

Fine-tune a YOLO detector for hard hats, measure mAP, and deploy a small model.

Computer vision · detectionAdvanced25 h
In developmentProject 8

Review sentiment: classic vs transformer

TF-IDF baseline against a fine-tuned DistilBERT, with error analysis of where each fails.

NLPIntermediate20 h
In developmentProject 9

Chat with your notes

Retrieval-augmented assistant with citations, an evaluation set, prompt-injection tests and a deployed API.

LLM · RAGIntermediate20 h
In developmentProject 10

Teach a small LLM a new skill

LoRA fine-tuning of a small open model for structured extraction, compared honestly with prompting alone.

LLM · fine-tuningAdvanced25 h

The standard every project follows

These are the habits that separate a portfolio notebook from work a team would trust in production.

Frame before you code
  • users, decision and timing
  • success measures and costs
  • what the model must not be used for
A real repository
  • src/ package, configs, tests, docs
  • pinned dependencies
  • Makefile: one command per step
Data you can trust
  • provenance and checksums
  • schema validation that stops bad data
  • named, counted cleaning rules
No leakage
  • split the way the model will be used
  • point-in-time features
  • preprocessing inside the pipeline
Honest modelling
  • baseline first
  • tuning with grouped cross-validation
  • test set used exactly once
Decisions, not just scores
  • calibrated probabilities
  • thresholds from business costs
  • confidence intervals
Responsible AI
  • fairness by group
  • explanations
  • a model card with limits
Engineering quality
  • type hints, lint, format
  • unit, behaviour and API tests
  • CI on every push
Shipping
  • validated API with versioned responses
  • Docker image
  • experiment tracking
After launch
  • drift monitoring
  • retraining and rollback plan
  • no personal data in logs

How to use a project

  1. Read the problem framing before writing any code.
  2. Work through the steps in order; each links back to the course topic it relies on.
  3. Write your own version first, then compare with the reference repository.
  4. Use each step's Check yourself box to confirm your results make sense.
  5. Finish with the portfolio step: a README, a model card and a short write-up of what you learned.