18 stations. 3 tracks. From statistics, Python and SQL to classic ML, feature engineering and deep learning, then experimentation, causal inference and doing data science in the LLM era — become the person who turns data into decisions.
18 stations on 3 ladders, in order — each rung is written assuming the one above it. Mark them off as you go; your progress is saved in this browser.
0 / 18
Pick up at
Track 1 of 3
FoundationsStart here
6 rungs · ~6h · ends at Data Wrangling at Scale.
0/6 done
1
Statistics & Probability
Start here
F1 · Beginner · 65 min
Statistics is the grammar of data science — misread it and every conclusion after is suspect. Learn distributions, sampling, expectation and variance, the central limit theorem, confidence intervals and hypothesis testing, plus the p-value everyone quotes and few define correctly.
Distributions & sampling
Central limit theorem
Confidence intervals
Hypothesis testing & p-values
Build itA test shows p = 0.03. Write, in one sentence, what that actually means — and what it does NOT mean.
✓ CheckpointExplain what a confidence interval does and does not claim — “95% chance the true value is in here” is the usual wrong version.
Python is the lingua franca of data work. Learn NumPy for vectorized math, pandas for wrangling, and the notebook workflow — the daily tools where 80% of a data scientist’s time is actually spent (cleaning, reshaping, joining), not on the glamorous modeling.
NumPy vectorization
pandas wrangling
Notebooks
Reshape, join, clean
Build itGiven a messy CSV with nulls and mixed types, produce a clean, typed dataframe ready to model. Which step is most likely to hide a bug?
✓ CheckpointExplain why vectorised NumPy beats a Python loop by enough to change what is feasible.
The data lives in databases, and SQL is how you get it. Learn joins, aggregation, window functions and CTEs well enough to answer real questions without exporting everything to pandas — because the query that runs in the warehouse beats the one that melts your laptop.
Joins & aggregation
Window functions
CTEs
Query performance
Build itCompute a 7-day rolling average of signups per country in SQL. Why is doing this in the warehouse better than pulling raw rows?
✓ CheckpointExplain what a window function does that GROUP BY cannot, using a running total as the example.
Before modeling comes looking. Learn to profile a dataset, spot missingness and outliers, understand distributions and correlations, and visualize honestly — because the model is only as good as your understanding of the data feeding it, and EDA is where you earn that understanding.
Data profiling
Missingness & outliers
Correlation vs causation
Honest visualization
Build itA feature correlates strongly with your target. List three ways that correlation could be misleading you before you trust it.
✓ CheckpointGive a real pair of variables that correlate with no causal link between them, and say what would distinguish the two cases.
The math under the models. Learn vectors and matrices, dot products and matrix multiplication, and gradients — just enough to understand what an algorithm is doing rather than treating it as a black box you feed and pray to. Intuition over proofs, but real intuition.
Vectors & matrices
Matrix multiplication
Gradients & derivatives
Geometric intuition
Build itExplain why gradient descent "walks downhill" using nothing but the idea of a derivative. What is the gradient pointing at?
✓ CheckpointExplain what a gradient tells you geometrically, and why that is enough to optimise anything.
Real data is bigger and messier than a tutorial CSV. Learn to handle data that does not fit in memory, join across sources, manage data types and encodings, and build reproducible pipelines — the unglamorous plumbing that determines whether your analysis is trustworthy.
Out-of-memory data
Multi-source joins
Encodings & types
Reproducible pipelines
Build itYour analysis gives different numbers each run. List the three most likely non-determinism sources and how to pin each down.
✓ CheckpointExplain what changes about your approach the moment the data stops fitting in memory.
The core of applied ML: learn from labeled examples. Learn linear and logistic regression, decision trees, the bias-variance trade-off, and train/validation/test discipline — the workflow and the failure modes (overfitting, leakage) that every model you build will live or die by.
Regression & classification
Trees & linear models
Bias-variance trade-off
Train/val/test discipline
Build itYour model scores 99% on training and 70% on test. Name the disease, then two ways to treat it.
✓ CheckpointState the bias-variance position of a model scoring 99% on train and 70% on test, and what you would change first.
Accuracy lies more often than it tells the truth. Learn precision, recall, F1, ROC/AUC, cross-validation, and the crucial skill of choosing the right metric for the problem — because optimizing accuracy on imbalanced data ships a model that predicts "no" and looks brilliant.
Precision, recall, F1
ROC/AUC
Cross-validation
Choosing the right metric
Build itA fraud model is 99.9% accurate. Explain why that could be worthless, and which metric you would actually report.
✓ CheckpointExplain why accuracy is the wrong metric on a 1%-positive dataset, and which one you would report instead.
Often the highest-leverage work in ML: the right features beat a fancier model. Learn encoding, scaling, handling missing values, creating interactions, and the discipline of preventing target leakage — the subtle bug that makes offline metrics soar and production crash.
Encoding & scaling
Handling missingness
Feature creation
Preventing leakage
Build itA feature secretly encodes the answer (leakage). Give a realistic example and explain how a time-based split would catch it.
✓ CheckpointExplain why fitting an encoder or scaler before the split leaks the test set, and how the leak shows up as an optimistic score.
The models that win on tabular data. Learn random forests and gradient boosting (XGBoost, LightGBM), why combining weak learners beats one strong one, and when a boosted-tree model — not a neural network — is still the correct, boring, effective answer for structured data.
Random forests
Gradient boosting (XGBoost)
Weak learners combined
Tabular > deep learning
Build itFor a tabular business dataset, argue why XGBoost is likely the right first model over a neural net. When would that flip?
✓ CheckpointExplain how boosting turns weak learners into a strong one, and why that makes it easier to overfit than bagging.
5
Unsupervised Learning
T5 · Advanced · 55 min
Finding structure with no labels. Learn clustering (k-means, hierarchical), dimensionality reduction (PCA, t-SNE/UMAP), and anomaly detection — the tools for segmenting customers, compressing features, and surfacing the weird points that supervised learning never labeled.
When the data is unstructured — images, text, audio — neural networks take over. Learn the building blocks (layers, activations, backprop), when deep learning is worth its cost over classic ML, and the frameworks (PyTorch) that turn the math into models you can actually train.
Neural network basics
Backprop & activations
When deep learning wins
PyTorch
Build itGiven a project, decide: classic ML or deep learning? List the signals in the data and problem that push you each way.
✓ CheckpointExplain when deep learning is worth it over a gradient-boosted tree on tabular data — the honest answer is “often not”.
The skill that most directly drives business decisions. Learn experiment design, sample-size and power calculations, avoiding peeking and p-hacking, and interpreting results honestly — because a badly run A/B test does worse than nothing: it launders a bad decision as data.
Experiment design
Power & sample size
Peeking & p-hacking
Honest interpretation
Build itA PM wants to stop the test early because it "looks significant". Explain the danger and what rule should govern stopping.
✓ CheckpointExplain what goes wrong when you check an experiment early and stop as soon as it looks significant.
Correlation is easy; causation is where decisions live and where careers are made. Learn confounders, the potential-outcomes framework, and quasi-experimental methods (difference-in-differences, instrumental variables) for when you cannot run an A/B test but still must answer "did it actually work?"
Confounders
Potential outcomes
Diff-in-diff & IV
Observational causality
Build itUsers who use feature X retain better. Design how you would tell "X causes retention" from "retained users use X".
✓ CheckpointName a confounder that would make an obvious causal story wrong, and how you would probe for it without an experiment.
A brilliant analysis nobody understands changes nothing. Learn to tell the story in the data, visualize for a decision (not for decoration), tailor depth to the audience, and state uncertainty honestly — the storytelling that turns a model into an action someone takes.
Data storytelling
Decision-focused viz
Audience tailoring
Communicating uncertainty
Build itPresent a churn model to executives in three slides. What is on them — and what technical detail did you deliberately leave off?
✓ CheckpointState the “so what” of a recent analysis in one sentence, before any numbers.
A model in a notebook helps no one. Learn to deploy models as services, monitor for data and concept drift, retrain on schedule, and version data and models — the engineering that turns a one-off analysis into a system that keeps making good predictions next quarter.
Model deployment
Drift monitoring
Retraining pipelines
Data & model versioning
Build itYour live model’s accuracy quietly drops over two months. Name the likely cause and the monitoring that would have caught it early.
✓ CheckpointExplain how you would notice a deployed model degrading before a stakeholder complains.
LLMs rewired the field. Learn where they help — text classification and extraction without labeled data, synthetic data, exploratory analysis by prompting — and where classic ML still wins, plus the honest evaluation that keeps a fluent-but-wrong model out of your decisions.
LLMs for classification/extraction
Zero-shot vs fine-tune
Synthetic data
Evaluating LLM outputs
Build itYou need to categorize 50K support tickets with no labels. Compare an LLM zero-shot approach vs training a classic model — costs and risks.
✓ CheckpointExplain when an LLM beats a trained classifier for an extraction task, and when it is the expensive wrong tool.
Data science splits into flavors — analytics, ML engineering, research, decision science — and knowing your lane matters. Learn the archetypes, the business acumen that separates senior from junior, how to pick high-impact problems, and how to grow from running analyses to shaping strategy.
DS archetypes
Business acumen
Picking high-impact work
IC to leadership
Build itMap your next six months: one stats/ML skill, one engineering skill, one communication skill — and the project that proves each.
✓ CheckpointName which data-science archetype your last six months resembled, and whether that is the one you want.
All 18 rungs on one screen. Each track assumes the one before it; the highlighted rung is where a track ends and the next one picks up.
Track order matters more than rung order inside a track: finish a track's last rung before starting the next track.
Worked example: a three-hours-a-week plan
Say you have three focused hours a week — an evening or two. Take the rungs in order and never split one across two weeks; a rung that would push the week past three hours starts the next week. The rung times add up to 17h 30m.
Week
Rungs
Time
Week 1
F1 Statistics & Probability, F2 Python for Data, F3 SQL & Data Access
2h 55m
Week 2
F4 Exploratory Data Analysis, F5 Linear Algebra & Calculus, F6 Data Wrangling at Scale
P5 Data Science with LLMs, P6 The Data Science Career
1h 45m
That is 7 weeks of reading. The Build-it exercise on each rung is extra and is where most of the learning happens, so plan to spend at least as long building as reading — or halve the pace and do both in the same week.
What each track costs in time
The rung times above, added up per track. Use it to plan the slow stretch: a track that takes a larger share of the total is where the plan above spends more weeks.
Track
Rungs
Time
Share
Longest rung
1. Foundations
6
5h 40m
32%
F1 Statistics & Probability (1h 5m)
2. Machine Learning
6
6h 5m
35%
T1 Supervised Learning (1h 5m)
3. Impact & the LLM Era
6
5h 45m
33%
P1 Experimentation & A/B Testing (1h 5m)
Common mistakes on the Data Scientist path
Optimizing accuracy on imbalanced data
When one class is rare, a model that always predicts the common class scores high accuracy and is useless. Choose precision, recall or a cost-weighted metric from what each kind of error costs (T2).
Target leakage
A feature built from information that only exists after the moment of prediction makes offline metrics soar and production collapse. For every feature, ask when its value would be known (T3), and split by time when the problem has time in it.
Reaching for deep learning on tables
On structured, tabular data, gradient-boosted trees (T4) are usually the strong default. Start with a simple baseline such as logistic regression (T1) so you know what the complex model is beating.
Rushing past statistics
Skimming F1 leads to misreading p-values. A p-value is not the probability that your hypothesis is true, and stopping an experiment the moment it looks significant (P1) makes it unreliable.
A leaderboard portfolio
Notebooks chasing a competition score show tuning, not judgment. Show a question tied to a decision, a baseline, honest evaluation, and a clear write-up of what to do next (P3).
Where this shows up: jobs and interviews
The day-to-day work
Most time goes to F2 to F4: getting data with SQL, cleaning it and exploring it. Then comes modeling or experiment analysis, and a lot of writing for decision-makers (P3). The mix depends on the flavor of the role (P6).
Technical screens
Expect probability and statistics questions (F1), SQL (F3) and Python with pandas (F2), plus ML concept questions: bias and variance, overfitting, metric choice and leakage (T1 to T3).
The case or take-home
You get a dataset and build a model or analyze an experiment. Reviewers look at validation discipline, metric choice and the write-up. Product cases ask how you would measure whether a feature worked (P1, P2).
Showing it in a portfolio
Two or three end-to-end projects: one experiment or causal analysis, one model evaluated honestly against a baseline, and perhaps one LLM classifier checked against a hand-labeled sample (P5).
Check yourself
Open a question only after you have answered it out loud. If one stumps you, the rung in brackets is where to go back to.
Fraud is 1 in 1,000 transactions, and your model scores 99.9% accuracy. Is it good?
Not necessarily: predicting "not fraud" every time also scores 99.9%. Look at precision and recall on the fraud class, and set the threshold from the cost of a missed fraud versus a false alarm. (T2)
A churn model is excellent offline and poor in production. One feature is support tickets in the last 30 days, computed when the training data was extracted. What went wrong?
Leakage. For past customers, that window includes tickets filed while they were already leaving. Compute every feature as of the prediction date, and validate on a time-based split. (T3)
A new onboarding flow launched in one country first, with no A/B test. How do you estimate its effect on retention?
Difference-in-differences against a comparable country, after checking their trends moved together before the launch. A simple before-and-after comparison would mix the effect with seasonality and other changes. (P2)
You must sort thousands of support tickets into a few categories and have no labels. LLM or trained classifier?
An LLM zero-shot is a fast start, but hand-label a sample to measure it. If cost or consistency matter at volume, use checked LLM labels to train a cheaper classic model. Either way, evaluate against the labeled sample. (P5)
Data Scientist roadmap — frequently asked questions
The common questions before you start — how long it takes, whether to follow it in order, and how it stays current.
How long does this roadmap take?
It runs 18 stations across 3 tracks — roughly ~17h of focused learning, plus the time you spend actually building. It is self-paced, so most people work through it over a few weeks, an evening or a single station at a time.
Do I have to follow the stations in order?
The tracks are ordered so each station builds on the one before, and following them start to finish is the intended path. But every station also stands alone — if you already have the foundations, jump straight to the part you need.
Is it free?
Yes. The whole roadmap, and every handbook, lab, and challenge it links to, is free and open — no sign-up and no paywall.
How is the roadmap kept current?
It teaches the durable fundamentals of the role first, then the tooling and the AI-era shifts on top — so most of it stays relevant as individual tools churn, and it is revised as the role itself changes.
Who is this roadmap for?
Anyone stepping into or leveling up in the Data Scientist role — whether you are switching in, early-career, or a senior filling gaps. Start where you are; the ladder shows what is left.