Stacc/Roadmaps/Data Science

Data Science learning roadmap

Frame a problem, build and validate a model, then deploy it and watch it in production.

9 modules About 204 hoursFree resources

Start with short learning sessions here. Continue coding and project work on your laptop with the same GitHub account.

Start this path
What you’ll build

Credit default risk model

A model that estimates default risk for existing credit-card accounts at a fixed point in time, and weighs the cost of each kind of mistake, served live from AWS. It is not for approving new loans.

Tap a milestone to see what you’ll build on your laptop.

  1. 1ML Fundamentals

    Write a one-page design doc and who uses the prediction, when it is made and what each mistake costs, then train two baseline pipelines.

    What to show: framing.md, two baseline pipelines and their tests.

  2. 3Feature Engineering

    Build preprocessing inside the pipeline, check for leakage and keep only the feature groups that earn their place.

    What to show: The feature pipeline, a leakage audit and an ablation table.

  3. 5Model Evaluation

    Pick a threshold by cost, check the probabilities are calibrated and compare errors across groups.

    What to show: evaluation.md with the threshold, calibration and group results.

  4. 7Advanced Models

    Compare a boosted tree model with your baseline on cost, speed and explainability, then decide whether to ship it.

    What to show: model_selection.md with the comparison, SHAP plots and the decision.

  5. 8Model Deployment

    Serve the model through a tested FastAPI app, run it in Docker locally, then put it live on AWS Lambda with a public URL.

    What to show: app/main.py, API tests, a Dockerfile and docs/api.md.

Tools: Model families, Segments, A/B design, Monitoring.

Credit recordsCredit recordsExisting card accounts with historical payment records. Given, not built.Existing card accounts with historical payment records.
03
FeaturesMilestone 03Feature EngineeringBuild preprocessing inside the pipeline, check for leakage and keep only the feature groups that earn their place.Reviewer looks for: The feature pipeline, a leakage audit and an ablation table.Leakage-safe, available at prediction time. Build preprocessing inside the pipeline, check for leakage and keep only the feature groups that earn their place.
01
BaselineMilestone 01ML FundamentalsWrite a one-page design doc and who uses the prediction, when it is made and what each mistake costs, then train two baseline pipelines.Reviewer looks for: framing.md, two baseline pipelines and their tests.Framed target, naive benchmark. Write a one-page design doc and who uses the prediction, when it is made and what each mistake costs, then train two baseline pipelines.
05
ChampionMilestone 05Model EvaluationPick a threshold by cost, check the probabilities are calibrated and compare errors across groups.Reviewer looks for: evaluation.md with the threshold, calibration and group results.Cost-aware threshold, calibrated, fairness-checked by group. Pick a threshold by cost, check the probabilities are calibrated and compare errors across groups.
07
ChallengerMilestone 07Advanced ModelsCompare a boosted tree model with your baseline on cost, speed and explainability, then decide whether to ship it.Reviewer looks for: model_selection.md with the comparison, SHAP plots and the decision.Gradient boosting; ships only if it earns its cost. Compare a boosted tree model with your baseline on cost, speed and explainability, then decide whether to ship it.
08
Model APIMilestone 08Model DeploymentServe the model through a tested FastAPI app, run it in Docker locally, then put it live on AWS Lambda with a public URL.Reviewer looks for: app/main.py, API tests, a Dockerfile and docs/api.md.Versioned, tested, with rollback. Serve the model through a tested FastAPI app, run it in Docker locally, then put it live on AWS Lambda with a public URL.
Risk reviewersRisk reviewersHuman review of account-risk monitoring. Given, not built.Human review of account-risk monitoring.

Runs on

02
Model familiesMilestone 02How Models LearnTrain five model families on the same folds, explain how each decides and pick a challenger to carry forward.Reviewer looks for: src/compare_models.py, tests that every model sees the same folds and reports/model_comparison.md.Five models compared on the same folds. Train five model families on the same folds, explain how each decides and pick a challenger to carry forward.
04
SegmentsMilestone 04Unsupervised LearningReduce the bill and payment columns with PCA, group accounts with k-means and test whether the segment helps the model.Reviewer looks for: src/segments.py, a leakage test and reports/segments.md.PCA and k-means account segments. Reduce the bill and payment columns with PCA, group accounts with k-means and test whether the segment helps the model.
06
A/B designMilestone 06A/B TestingPlan a test of a reminder for at-risk accounts: hypothesis, sample size, stopping rule and ship rule.Reviewer looks for: experiment_spec.md and a seeded simulation of peeking.How the intervention gets tested. Plan a test of a reminder for at-risk accounts: hypothesis, sample size, stopping rule and ship rule.
09
MonitoringMilestone 09Data Science CapstoneRelease the model live with green CI, a model card, one final test, an updated design doc, review rules and a decision memo.Reviewer looks for: MODEL_CARD.md, the final test report, the decision memo and a walkthrough link.Model card, fairness review, alerts. Release the model live with green CI, a model card, one final test, an updated design doc, review rules and a decision memo.
Hover a part to see what you build. Numbers are the milestones; dashed tiles are the data you start with and the people who use the result. Scroll sideways to see the whole system.

What you'll build

Credit default risk model

A model that estimates default risk for existing credit-card accounts at a fixed point in time, and weighs the cost of each kind of mistake, served live from AWS. It is not for approving new loans.

Data

Default of Credit Card Clients (UCI, CC BY 4.0): 30,000 existing accounts with April–September 2005 payment/bill history and a default flag. State the prediction time and feature availability. Error costs are teaching assumptions in declared units.

  • Default of Credit Card Clients (UCI, CC BY 4.0)

One GitHub repo for the whole path. You connect it once, and each module adds a verified milestone.

You need: A laptop with 8 GB of RAM, Docker from Model Deployment onward, and an AWS account on the Free plan to put the model live. If AWS will not accept your card, a GitHub-only route gets you through the same checks.

Tap a milestone to see what you’ll build on your laptop.

  1. 1ML Fundamentals

    Write a one-page design doc and who uses the prediction, when it is made and what each mistake costs, then train two baseline pipelines.

    What to show: framing.md, two baseline pipelines and their tests.

  2. 3Feature Engineering

    Build preprocessing inside the pipeline, check for leakage and keep only the feature groups that earn their place.

    What to show: The feature pipeline, a leakage audit and an ablation table.

  3. 5Model Evaluation

    Pick a threshold by cost, check the probabilities are calibrated and compare errors across groups.

    What to show: evaluation.md with the threshold, calibration and group results.

  4. 7Advanced Models

    Compare a boosted tree model with your baseline on cost, speed and explainability, then decide whether to ship it.

    What to show: model_selection.md with the comparison, SHAP plots and the decision.

  5. 8Model Deployment

    Serve the model through a tested FastAPI app, run it in Docker locally, then put it live on AWS Lambda with a public URL.

    What to show: app/main.py, API tests, a Dockerfile and docs/api.md.

Tools: Model families, Segments, A/B design, Monitoring.

Credit recordsCredit recordsExisting card accounts with historical payment records. Given, not built.Existing card accounts with historical payment records.
03
FeaturesMilestone 03Feature EngineeringBuild preprocessing inside the pipeline, check for leakage and keep only the feature groups that earn their place.Reviewer looks for: The feature pipeline, a leakage audit and an ablation table.Leakage-safe, available at prediction time. Build preprocessing inside the pipeline, check for leakage and keep only the feature groups that earn their place.
01
BaselineMilestone 01ML FundamentalsWrite a one-page design doc and who uses the prediction, when it is made and what each mistake costs, then train two baseline pipelines.Reviewer looks for: framing.md, two baseline pipelines and their tests.Framed target, naive benchmark. Write a one-page design doc and who uses the prediction, when it is made and what each mistake costs, then train two baseline pipelines.
05
ChampionMilestone 05Model EvaluationPick a threshold by cost, check the probabilities are calibrated and compare errors across groups.Reviewer looks for: evaluation.md with the threshold, calibration and group results.Cost-aware threshold, calibrated, fairness-checked by group. Pick a threshold by cost, check the probabilities are calibrated and compare errors across groups.
07
ChallengerMilestone 07Advanced ModelsCompare a boosted tree model with your baseline on cost, speed and explainability, then decide whether to ship it.Reviewer looks for: model_selection.md with the comparison, SHAP plots and the decision.Gradient boosting; ships only if it earns its cost. Compare a boosted tree model with your baseline on cost, speed and explainability, then decide whether to ship it.
08
Model APIMilestone 08Model DeploymentServe the model through a tested FastAPI app, run it in Docker locally, then put it live on AWS Lambda with a public URL.Reviewer looks for: app/main.py, API tests, a Dockerfile and docs/api.md.Versioned, tested, with rollback. Serve the model through a tested FastAPI app, run it in Docker locally, then put it live on AWS Lambda with a public URL.
Risk reviewersRisk reviewersHuman review of account-risk monitoring. Given, not built.Human review of account-risk monitoring.

Runs on

02
Model familiesMilestone 02How Models LearnTrain five model families on the same folds, explain how each decides and pick a challenger to carry forward.Reviewer looks for: src/compare_models.py, tests that every model sees the same folds and reports/model_comparison.md.Five models compared on the same folds. Train five model families on the same folds, explain how each decides and pick a challenger to carry forward.
04
SegmentsMilestone 04Unsupervised LearningReduce the bill and payment columns with PCA, group accounts with k-means and test whether the segment helps the model.Reviewer looks for: src/segments.py, a leakage test and reports/segments.md.PCA and k-means account segments. Reduce the bill and payment columns with PCA, group accounts with k-means and test whether the segment helps the model.
06
A/B designMilestone 06A/B TestingPlan a test of a reminder for at-risk accounts: hypothesis, sample size, stopping rule and ship rule.Reviewer looks for: experiment_spec.md and a seeded simulation of peeking.How the intervention gets tested. Plan a test of a reminder for at-risk accounts: hypothesis, sample size, stopping rule and ship rule.
09
MonitoringMilestone 09Data Science CapstoneRelease the model live with green CI, a model card, one final test, an updated design doc, review rules and a decision memo.Reviewer looks for: MODEL_CARD.md, the final test report, the decision memo and a walkthrough link.Model card, fairness review, alerts. Release the model live with green CI, a model card, one final test, an updated design doc, review rules and a decision memo.
Hover a part to see what you build. Numbers are the milestones; dashed tiles are the data you start with and the people who use the result. Scroll sideways to see the whole system.
Final demo question
How does your model score an account, why did you pick that threshold, where is it less sure, and can someone else reproduce it? Explain why it cannot be used to approve new loans.

Module by module

  1. 1ML Fundamentals

    Frame what you are predicting and what a wrong guess costs, then build a simple baseline model.

    Write a one-page design doc and who uses the prediction, when it is made and what each mistake costs, then train two baseline pipelines.

    Frame the prediction · Baselines and the right metric · Pipelines and a clean test set

    01

    ML Fundamentals

    Write a one-page design doc and who uses the prediction, when it is made and what each mistake costs, then train two baseline pipelines.

    23h
  2. 2How Models Learn

    See inside the classic models: regression and its checks, logistic regression, naive Bayes, KNN, trees, forests and regularisation.

    Train five model families on the same folds, explain how each decides and pick a challenger to carry forward.

    Regression and its checks · Classifiers: logistic, naive Bayes, KNN · Trees, forests and regularisation

    02

    How Models Learn

    Train five model families on the same folds, explain how each decides and pick a challenger to carry forward.

    24h
  3. 3Feature Engineering

    Turn raw columns into useful features without leaking the answer, and keep only the ones that help.

    Build preprocessing inside the pipeline, check for leakage and keep only the feature groups that earn their place.

    Encoding and scaling · Avoiding data leakage · Keeping features that help

    03

    Feature Engineering

    Build preprocessing inside the pipeline, check for leakage and keep only the feature groups that earn their place.

    18h
  4. 4Unsupervised Learning

    Find structure without labels: reduce columns with PCA, group accounts into segments and test whether segments help the model.

    Reduce the bill and payment columns with PCA, group accounts with k-means and test whether the segment helps the model.

    PCA · k-means · Hierarchical clustering and segments

    04

    Unsupervised Learning

    Reduce the bill and payment columns with PCA, group accounts with k-means and test whether the segment helps the model.

    18h
  5. 5Model Evaluation

    Pick a decision threshold, check the probabilities are honest and see who the model gets wrong.

    Pick a threshold by cost, check the probabilities are calibrated and compare errors across groups.

    Choosing a threshold · Honest probabilities · Errors by group

    05

    Model Evaluation

    Pick a threshold by cost, check the probabilities are calibrated and compare errors across groups.

    22h
  6. 6A/B Testing

    Plan an experiment big enough to answer the question, decide the rules up front and read the result honestly.

    Plan a test of a reminder for at-risk accounts: hypothesis, sample size, stopping rule and ship rule.

    Designing the test · Sample size · When to stop and decide

    06

    A/B Testing

    Plan a test of a reminder for at-risk accounts: hypothesis, sample size, stopping rule and ship rule.

    20h
  7. 7Advanced Models

    Try a stronger model like XGBoost, explain it with SHAP and ship it only if it is worth the extra cost.

    Compare a boosted tree model with your baseline on cost, speed and explainability, then decide whether to ship it.

    Boosted trees · Validating it properly · Explaining it and the trade-off

    07

    Advanced Models

    Compare a boosted tree model with your baseline on cost, speed and explainability, then decide whether to ship it.

    24h
  8. 8Model Deployment

    Serve your model through a tested API that checks its inputs, run it in Docker, then put it live on AWS.

    Serve the model through a tested FastAPI app, run it in Docker locally, then put it live on AWS Lambda with a public URL.

    A model API with FastAPI · Testing the API · Docker, then live on AWS

    08

    Model Deployment

    Serve the model through a tested FastAPI app, run it in Docker locally, then put it live on AWS Lambda with a public URL.

    27h
  9. 9Data Science Capstone

    Ship the full project with a model card, one honest final test, human review rules and a decision memo.

    Release the model live with green CI, a model card, one final test, an updated design doc, review rules and a decision memo.

    The model card · The final test · Human review and the memo

    09

    Data Science Capstone

    Release the model live with green CI, a model card, one final test, an updated design doc, review rules and a decision memo.

    28h

More roadmaps

  • Foundations roadmapPython, SQL, Git and statistics. The base every data role builds on.
  • Data Engineering roadmapBuild one data platform end to end: ingest, transform, schedule, deploy and monitor it.
  • Data Analysis roadmapTurn messy data and vague questions into analysis people trust and act on.
  • MLOps roadmapTake the model you built in Data Science to production: test it, ship it, monitor it, and improve it safely.

Turn this map into shipped work.

Finish with Credit default risk model on your GitHub.

Start this path