Przejdź do treści

Główna treść tej strony jest po angielsku.Przetłumacz tę stronę

Ścieżki kariery

Data Scientist roadmap

Data scientists combine statistics, programming and business sense to predict outcomes and find insights. Today the role also includes using LLMs for text data and analysis.

Czas
⏱ 6–8 months
Level
Intermediate
Steps
6 steps
Architecture ↓Cheat sheet ↓5 interview Q&As ↓
Udostępnij

The roadmap

Click a step to open the full guide.

Best architecture: a data science workflow in production

How analysis and models move from notebooks to something the business uses every day.

From question to impact

  1. Business question
  2. Warehouse data (SQL)
  3. Notebook exploration
  4. Model + validation
  5. Scheduled job / API / app
  6. Dashboard + monitoring

Why it's built this way

  • Pull data with SQL from the warehouse, not ad-hoc exports.
  • Turn notebooks into versioned, scheduled code once results matter.
  • Ship results where people work: dashboards, apps or APIs.
  • Monitor model accuracy and data quality over time.

Data Scientist cheat sheet

The essentials on one screen. Bookmark it for revision.

Which method?

Predict a numberRegression (linear, boosting)
Predict a categoryClassification (logistic, boosting)
Find groupsClustering (k-means, HDBSCAN)
Did the change work?A/B test / causal methods
ForecastTime-series models
Text themesEmbeddings + clustering / LLMs

pandas essentials

Missing valuesdf.isna().sum()
Value countsdf["col"].value_counts()
Mergedf.merge(other, on="id", how="left")
Groupdf.groupby("g").agg(total=("x", "sum"))
Datespd.to_datetime(df["date"])

Improve performance

  • Sample large data for exploration; run final analysis on full data.
  • Push heavy aggregation into SQL / the warehouse.
  • Use vectorised pandas or faster libraries (Polars) for big data.
  • Cache intermediate results to files (Parquet) instead of recomputing.

Security checklist

  • Anonymise or pseudonymise personal data before analysis.
  • Follow data-access rules; don't copy production data to personal devices.
  • Check models for unfair bias across groups.
  • Don't send confidential data to external AI tools without approval.

Most important interview questions

Click a question to see a strong answer. Each step also has its own questions.

1Czym jest the difference between a data analyst and a data scientist?

Analysts focus on describing what happened and why, with SQL and dashboards. Data scientists also predict and experiment, using statistics and machine learning.

2How do you handle an imbalanced dataset?

Use appropriate metrics (PR-AUC, recall), class weights or resampling, tune the threshold, and gather more minority data if possible.

3Czym jest overfitting and how do you prevent it?

Learning noise instead of patterns. Prevent with validation, simpler models, regularisation, more data and early stopping.

4Design an A/B test for a new checkout button.

Define the metric (conversion), compute sample size for the minimum effect you care about, randomise users, run for full weeks, check for issues, then test significance and practical impact.

5How do you explain a model to non-technical stakeholders?

Focus on what it predicts, how accurate it is in business terms, the main drivers (SHAP), limitations, and how they should use it.

Related map drogowych

6

Data Analyst

6 steps · 3–4 months · Beginner

Turn data into answers and dashboards.

View roadmap →
6

Machine Learning Engineer

6 steps · 6–9 months · Advanced

Train, deploy and monitor machine learning models.

View roadmap →
6

Python

6 steps · 6–8 weeks · Beginner

The language of AI, data and automation.

View roadmap →
1

Data Scientist · Step 1 of 6 · 4 weeks

Python, pandas and SQL

Your daily tools for working with data.

Data scientists spend most of their time getting and cleaning data. Fluency in pandas and SQL makes everything else faster.

Learn to write reusable functions instead of copying cells around notebooks.

What you'll learn

  • Python basics
  • pandas and NumPy
  • SQL with joins and window functions
  • Jupyter notebooks
  • Data cleaning
  • Git

Data science workflow

  1. Question
  2. Get data (SQL)
  3. Clean (pandas)
  4. Odkryj
  5. Model
  6. Communicate

Key concepts

Tidy data
Each row an observation, each column a variable.
merge
pandas' SQL JOIN.

Step by step

  1. 1.1

    pandas fluency

    Filter, join (merge), group, reshape, handle dates.

  2. 1.2

    SQL fluency

    Joins, CTEs, window functions on real tables.

  3. 1.3

    Reusable code

    Functions and modules instead of copy-pasted notebook cells.

Notes

  • Clean data beats a clever model.

Tips & tricks

  • Practise on messy real data, not only tidy course datasets.

Interview questions

How do you join two DataFrames in pandas?

pd.merge(left, right, on="id", how="left"); check row counts before and after to catch duplicate keys.

The quick way to study this

Two weeks of Python + pandas, two weeks of daily SQL practice.

Python roadmap →
2

Data Scientist · Step 2 of 6 · 3 weeks

Statistics and experiments

Reason correctly under uncertainty.

Statistics is what makes a data scientist trustworthy. Know how to design an experiment, check whether a result is real and explain uncertainty.

Causal thinking (does X cause Y?) is increasingly valued.

What you'll learn

  • Probability and distributions
  • Hypothesis testing
  • Confidence intervals
  • A/B test design
  • Regression analysis
  • Bias and confounders

Hypothesis testing

  1. Null hypothesis
  2. Collect data
  3. Test statistic
  4. p-value
  5. Reject or not

Key concepts

p-value
The chance of seeing results at least this extreme if there were no real effect.
Confidence interval
A range likely to contain the true value.
Confounder
A hidden factor affecting both cause and effect.

Step by step

  1. 2.1

    Distributions

    Normal, binomial, skewed data; the central limit theorem.

  2. 2.2

    Hypothesis tests

    t-tests, chi-square; p-values and confidence intervals.

  3. 2.3

    Experiments

    A/B test design, power and sample size.

  4. 2.4

    Causality basics

    Confounders, randomisation, difference-in-differences.

Notes

  • Correlation is not causation; look for confounders.

Tips & tricks

  • Explain p-values in plain words; interviewers ask.

Interview questions

Explain a p-value to a manager.

If the change really did nothing, a p-value of 0.03 means we'd see a result this big only about 3% of the time by chance. It's evidence, not proof, and says nothing about how big or important the effect is.

The quick way to study this

Learn each concept with a simulation in Python; it builds intuition fast.

3

Data Scientist · Step 3 of 6 · 2 weeks

Exploratory analysis and visualisation

Find the story in the data.

Exploratory data analysis (EDA) is where you find problems in data and the first insights. Always explore before modelling.

Good charts persuade decision-makers more than accuracy numbers.

What you'll learn

  • Distributions and outliers
  • Porównania: Grouped
  • Time series plots
  • Correlation analysis
  • Clear chart design

EDA checklist

  • EDA
  • Data quality: nulls, duplicates, types
  • Distributions: histograms, outliers
  • Relationships: correlations, scatter
  • Segments: by region, product, time
  • Hypotheses to test

Key concepts

EDA
Exploratory data analysis: understand the data before modelling.
Outlier
A value far from the rest; investigate before removing.

Step by step

  1. 3.1

    Profile

    Shape, types, missing values, unique counts.

  2. 3.2

    Visualise

    Histograms, box plots, scatter plots, time series.

  3. 3.3

    Segment

    Compare groups; look for surprises.

Notes

  • Look at raw rows, not just summaries.

Tips & tricks

  • Write down hypotheses before you explore, then test them.

Interview questions

What do you do first with a new dataset?

Understand where it came from and what each column means, check quality (missing, duplicates, ranges), look at distributions, then form hypotheses.

The quick way to study this

Do one full EDA notebook on a dataset you care about.

4

Data Scientist · Step 4 of 6 · 5 weeks

Machine learning

Build, evaluate and explain predictive models.

Learn the scikit-learn workflow deeply. Gradient-boosted trees are usually the strongest model for business tables.

Explainability matters: stakeholders want to know why the model predicts something.

What you'll learn

  • Regression and classification
  • Tree-based models and boosting
  • Feature engineering
  • Cross-validation and metrics
  • Model explainability (SHAP)
  • Clustering

Modelling loop

  1. Baseline
  2. Features
  3. Train
  4. Cross-validate
  5. Error analysis
  6. Improve

Key concepts

Feature engineering
Creating inputs that help the model (e.g. days since last purchase).
SHAP
Explains how much each feature pushed a prediction up or down.

Step by step

  1. 4.1

    Frame the problem

    Classification, regression, ranking or clustering?

  2. 4.2

    Baseline then models

    Simple rule → logistic/linear → boosted trees.

  3. 4.3

    Validate properly

    Cross-validation; time-based splits for time data.

  4. 4.4

    Explain

    Feature importance and SHAP for stakeholders.

Notes

  • Pick the metric based on the business cost of errors.

Tips & tricks

  • Always beat a simple baseline before trying complex models.

Interview questions

How would you build a churn model?

Define churn clearly, build features from behaviour (recency, frequency, tickets), use a time-based split, start with logistic regression then boosting, measure PR-AUC/recall at a useful threshold, and explain drivers with SHAP.

The quick way to study this

Complete three end-to-end modelling projects with different problem types.

ML Engineer roadmap →
5

Data Scientist · Step 5 of 6 · 2 weeks

LLMs and text data

Use modern AI for text analysis.

Reviews, tickets and survey answers are now easy to analyse with LLMs and embeddings. This is a fast-growing part of data science work.

Always validate LLM labels against a human-labelled sample.

What you'll learn

  • Embeddings
  • Text classification with LLMs
  • Topic discovery
  • Sentiment and extraction
  • Evaluating LLM outputs

Analysing 10,000 reviews

  1. Reviews
  2. Embed
  3. Cluster
  4. LLM names themes
  5. Validate sample
  6. Report

Key concepts

Zero-shot classification
An LLM labels text without training examples.
Label quality
Agreement between LLM labels and human labels.

Step by step

  1. 5.1

    Embeddings

    Turn text into vectors for clustering and search.

  2. 5.2

    LLM labelling

    Classify text with an LLM; check against human labels.

  3. 5.3

    Topic discovery

    Cluster embeddings and name the clusters.

Notes

  • LLM labels need checking like any other measurement.

Tips & tricks

  • Use embeddings + clustering to discover themes in thousands of reviews.

Interview questions

How do you validate LLM-generated labels?

Label a random sample by hand, compare (accuracy, confusion matrix, agreement), refine the prompt or categories, and report the measured accuracy with the results.

The quick way to study this

Classify 1,000 product reviews with an LLM and check accuracy on 100 manual labels.

6

Data Scientist · Step 6 of 6 · 2 weeks

Communication and portfolio

Turn analysis into decisions.

The best data scientists change decisions. Practise writing a one-paragraph summary with the recommendation first.

A simple deployed app makes your project memorable.

What you'll learn

  • Writing executive summaries
  • Presenting results
  • Deploying a simple app (Streamlit)
  • Case-study portfolio

A strong results memo

  1. Recommendation
  2. Key evidence (3 points)
  3. Impact estimate
  4. Risks & next steps
  5. Appendix: method

Key concepts

Pyramid principle
Lead with the answer, then the supporting arguments.

Step by step

  1. 6.1

    Executive summary

    Recommendation first, then 3 supporting points.

  2. 6.2

    Interactive demo

    A small Streamlit app for your best model.

  3. 6.3

    Portfolio

    3 projects with clear business impact.

Notes

  • Lead with the recommendation, then the evidence.

Tips & tricks

  • Build a Streamlit app for your best model so anyone can try it.

Interview questions

Tell me about a time your analysis changed a decision.

Use STAR: the decision at stake, your analysis, how you communicated it, and the measurable result.

The quick way to study this

Rewrite your best project as a one-page business memo.

Udostępnij tę stronę