Machine Learning
Topic 40 ยท Job ready ยท โฑ about 60 h

Portfolio Projects: Build, Publish, Present

Two guided end-to-end projects on real public data, a repository template, a README that gets read, and how to present your work.

Show:
In this topic
Beginner

What a hiring manager looks for

A portfolio project is evidence that you can do the job. Reviewers usually spend two or three minutes on it, so the result must be obvious from the top of the README. Strong projects share five qualities:

  1. A real question with a decision behind it ("which bikes should be moved overnight?"), not "I applied 7 algorithms to Titanic".
  2. Real, messy data that you cleaned yourself, with the cleaning explained.
  3. Honest evaluation: a baseline, a split that matches deployment (time-based if relevant), a metric that matches the decision, error analysis and limitations.
  4. Clean engineering: a repository someone else can run (requirements, clear structure, a few tests), not one giant notebook.
  5. Communication: a README that leads with the result and a short write-up or demo.

Two good projects beat ten tutorial copies. If you have time for a third, make it in the domain you want to work in, or use an LLM/RAG system with an evaluation set.

Beginner

Project 1: bike-share demand forecasting

Question: how many bikes will be rented in each of the next 24 hours, so the operator can rebalance stations overnight? Data: the public UCI "Bike Sharing" hourly dataset (two years of hourly rentals with weather and calendar), available in scikit-learn via OpenML. Skills shown: time-series validation, lag and calendar features, baselines, gradient boosting, error analysis by hour and weather, uncertainty with quantile models.

Milestones: (1) explore and plot daily and weekly patterns; (2) seasonal-naive baseline; (3) features with no leakage; (4) gradient boosting with a rolling-origin backtest; (5) prediction intervals with quantile loss; (6) error analysis; (7) README with a chart and the headline number (e.g. "x% lower MAE than the seasonal baseline"). The notebook below guides every step in Google Colab.

Beginner

Project 2: income prediction with a fairness audit

Question: can census attributes predict whether income exceeds $50k, and does the model perform equally well across groups? Data: the public UCI "Adult" census dataset (about 48,000 people), via OpenML. Skills shown: cleaning categorical data, a scikit-learn Pipeline with a ColumnTransformer, cross-validated model comparison, threshold selection, calibration, per-group metrics, SHAP-style explanations with permutation importance, and a model card.

This is deliberately a project about responsible ML: the dataset contains sex and race, and the interesting result is not the accuracy but the gaps between groups and what you would (and would not) do with such a model. Write that up honestly; interviewers notice.

Try it

The guided project notebooks

Open them in Google Colab (File โ†’ Upload notebook, or copy the code). They download the datasets themselves; each step has a task, hints and a reference solution you can reveal after trying.

Interactive: this part runs in your browser.

Beginner

Repository template and README

Publish each project as its own GitHub repository with this structure:

bike-demand-forecast/ โ”œโ”€โ”€ README.md โ† the most important file โ”œโ”€โ”€ requirements.txt โ”œโ”€โ”€ .gitignore (data/, .venv/, *.joblib) โ”œโ”€โ”€ notebooks/ โ”‚ โ””โ”€โ”€ 01_exploration.ipynb โ”œโ”€โ”€ src/bikes/ โ”‚ โ”œโ”€โ”€ data.py load + clean โ”‚ โ”œโ”€โ”€ features.py lag + calendar features โ”‚ โ””โ”€โ”€ train.py train, evaluate, save โ”œโ”€โ”€ tests/test_features.py e.g. "no feature uses future values" โ””โ”€โ”€ reports/figures/ charts used in the README

A README that gets read, in this order:

# Bike-share demand forecasting (24 h ahead) **Result:** gradient boosting with lag and weather features cuts hourly MAE by 28% versus a seasonal-naive baseline (rolling backtest over 8 weeks). [chart] ## Problem who uses the forecast, and for what decision ## Data source, size, period, cleaning steps, known issues ## Approach baseline โ†’ features โ†’ model โ†’ validation (and why) ## Results table of metrics vs baselines; error analysis by hour/weather ## Limitations what it can't do; what you'd do next ## Reproduce git clone โ€ฆ ยท pip install -r requirements.txt ยท python -m bikes.train
Try it

Try it: score your project

The rubric reviewers use, roughly. Tick what your project already has; aim for at least 80% before you share it.

Interactive: this part runs in your browser.

Beginner

Presenting your work

  • Pin your best two repositories on your GitHub profile, with a one-line description each.
  • Write a short blog post or LinkedIn article per project: the problem, one chart, the result, one thing that surprised you. It shows communication and gets shared.
  • Prepare a two-minute spoken version of each project for interviews: problem โ†’ data โ†’ approach โ†’ result with a number โ†’ what you'd improve.
  • Expect follow-up questions: "Why that metric?", "How did you avoid leakage?", "What would you do with more time?", "How would you deploy it?". Practise answering them out loud.
  • Kaggle competitions and open-source contributions (even documentation fixes to libraries you use) are good additions, but they don't replace a project you designed end to end.
Check yourself

Quiz

1. What should the first lines of a project README show?

2. Which project is more impressive to a hiring manager?

3. For a forecasting project, which validation is credible?

4. Why include a few tests in a portfolio repo?

Go deeper

๐Ÿ“š Further reading

Trusted, mostly free sources: official documentation, well-known textbooks and courses. When you want to confirm or go deeper on anything here, start with these.

Ready