Sefism Data Science Roadmap

What this roadmap is

The data science roadmap is Sefism's ordered path for students and early-career engineers who want to become data scientists and feel lost in long courses with no clear order. Tauseef Fayyaz wrote it around the parts of the job that courses tend to skip: cleaning genuinely messy data, explaining statistics in plain words, running experiments honestly and choosing metrics that match a real decision.

You choose a focus at the start, and the roadmap adjusts its topics, resources and projects:

• Business data science (the default): analysis, experiments and models for decisions, in Python.
• Towards ML engineering: more modelling depth, such as hyperparameter tuning and a project taken towards production, still in Python.
• Statistics and R: research and statistics first, with the tidyverse, tidymodels, regression for inference and survey microdata.

By the end you should be able to:

• Load, clean, reshape and summarise a messy real dataset with pandas or the tidyverse, and keep a log of every cleaning decision.
• Choose the right statistical test or confidence interval for a question, and explain what the result does and does not mean.
• Design an A/B test with a sample size worked out in advance, check it is healthy and read the result without peeking or cherry-picking segments.
• Build a model with a proper split, baseline, pipeline and cross-validation, choose the metric that matches the decision, and find leakage before it fools you.
• Publish two or three question-driven projects with clear write-ups and rerunnable code, and know which first roles to apply for.

Who it suits, and who should start elsewhere

It suits students in computer science, statistics, economics or engineering who enjoy asking questions of data, and analysts who want to add statistics, experiments and modelling. It is honest about the boundary with neighbouring fields: the roadmap stops at modelling that a data scientist needs and sends you elsewhere for more.

If you want neural networks, deep learning and production ML, the Machine learning roadmap is the next step. If dashboards and reporting for business teams are what you enjoy, the Data analytics roadmap goes much deeper there. If you would rather build the pipelines than analyse the data, see Data engineering. And if you have never programmed, start with Programming foundations.

Prerequisites and realistic time

Some programming in Python (or a willingness to learn R from scratch on the R focus), school-level maths, and a laptop that runs notebooks. Google Colab works if your laptop is slow. The roadmap adds up to about 215 hours across 10 stages:

• At about 5 hours a week (the default on the public page): roughly 43 weeks.
• At about 8 hours a week: roughly 31 weeks.
• At about 12 hours a week: roughly 24 weeks.

The first stage, on Python or R for data, is a foundation stage that is skipped or shortened if you already know it.

A sample week for a student (10 to 12 hours). Two evenings on the stage topics with StatQuest or the stage's course, one evening on the checklist, and a weekend block on the milestone notebook with your own dataset.

A sample week for someone working full time (5 hours). Two short weekday sessions on one concept each, explained out loud afterwards, and a three-hour weekend block in a notebook. Pick a dataset you care about in week one and keep using it, as the roadmap advises.

Stage by stage

1. Python for data: NumPy and pandas (25 hours, foundation)

Notebooks and a clean environment, NumPy arrays, loading and inspecting a DataFrame, selecting and filtering, grouping, joining and reshaping. On the R focus the stage becomes R for data: the tidyverse, with vectors and NA, tibbles and the dplyr verbs. Milestone: a notebook that loads a real public dataset, filters it to one clear population, joins a second table with a row-count check, reshapes it and prints three summary tables, rerunnable from a clean start. The advice: about twenty operations cover most real work.

2. Statistics and probability you actually use (30 hours)

Describing data, distributions, conditional probability and base rates, samples, the central limit theorem and standard error, confidence intervals, hypothesis tests and p-values, choosing the right test, and correlation, confounders and causation. Milestone: a notebook that describes three columns honestly, computes two intervals, runs two tests chosen from the question and explains each in two plain sentences.

3. SQL for data work (15 hours)

Aggregates with GROUP BY, joins at the right grain, window functions, CTEs, cohort and retention queries, and SQL from your notebook. Milestone: a saved set of analytical queries including a monthly cohort retention table, each run from your notebook into a data frame. SQL is often the first screen for data roles, and the skill people from Kaggle underrate most.

4. Data cleaning and EDA (20 hours)

Tidy shape, types, dates and categories, missing values, duplicates and keys, outliers, a first pass of exploratory analysis and a cleaning log. Milestone: a messy public dataset cleaned entirely by code, with a log of every change and row counts, and at least eight written findings and open questions.

5. Visualisation and storytelling (15 hours)

Choosing the right chart, plotting in code, showing distributions and uncertainty, action titles and writing up an analysis. Milestone: a one-page write-up with the answer in the first two lines, three charts, a caveats paragraph and one recommendation, tested on a non-technical reader.

6. Machine learning with scikit-learn (30 hours)

Framing a supervised problem, splitting, baselines, linear and logistic regression, random forests and gradient boosting, pipelines and cross-validation, with hyperparameter tuning on the ML focus. On the R focus the stage becomes Machine learning with tidymodels, adding modelling for inference versus prediction. Milestone: a modelling notebook with a time-aware or stratified split, a baseline, two models in pipelines, cross-validated scores with their spread and one final test result.

7. Experiments and A/B testing (20 hours)

Designing a test, the unit of randomisation, sample size and power, analysing the result, peeking, health checks with SRM and guardrails, multiple comparisons and practical versus statistical significance. Milestone: a full experiment report on a public A/B test dataset ending in a clear recommendation. Tauseef notes that experiments set data scientists apart and that many product companies screen for them.

8. Feature engineering and model evaluation (20 hours)

Creating features, encoding and scaling, finding leakage, the confusion matrix, precision and recall, ROC and precision-recall curves, regression metrics, thresholds and calibration, and imbalanced classes. Milestone: your earlier model improved with five documented features, a written leakage check and a threshold justified by costs or capacity.

9. Working with messy real data (15 hours)

Collecting from APIs, scraping responsibly, public and government data, documenting your data, sampling and selection bias, personal data and privacy, and survey microdata on the R focus. Milestone: a dataset you collected yourself, with raw responses saved, a datasheet and data dictionary, and a note on bias and privacy. Local data on prices, jobs, transport, education or health is suggested.

10. Portfolio, Kaggle profile and first role (25 hours)

Question-driven projects, the README and write-up, reproducibility, using Kaggle well, data roles and routes in, take-home analysis tasks, and, depending on focus, taking a project towards ML engineering or producing research outputs and replication packages. Milestone: two or three projects that rerun from scratch, one on data you collected, a Kaggle profile with one well-explained public notebook, and applications sent.

Projects to build alongside

• A/B Test Analysis Report on a Public Experiment Dataset, for every focus.
• Early Warning System for Student Disengagement, where leakage and thresholds really matter.
• Neighbourhood House Price Index from Open Property Sales Data, on the business and R focuses.
• Reproducing a Published Machine Learning Result with an Open Report, on the ML focus.
• Household Survey Cleaning and Weighting Toolkit for NGOs, on the R focus.

How it works on Sefism

Everything on Sefism is free to learn from. The public roadmap page lists all ten stages with their summaries and length, and opens the first two stages in full. Add the roadmap to your dashboard to open every stage, tick them off as you go and get the companion cheat sheet. Your first roadmap is free to add; after that, adding a roadmap uses credits, and new members get starting credits when they join.

In the dashboard every topic has a short explanation, a check question and an AI prompt that makes an assistant quiz you rather than do the analysis for you, and every stage has Tauseef's advice, a checklist and curated resources. You can ask Tauseef a question about any stage.

The companion cheat sheet

The Data science cheat sheet revises every stage in ten chapters, from data frames and statistics to experiments, evaluation and the first role, with an edition for each focus.

Related on Sefism

• Difference between data analyst, data scientist and machine learning engineer
• Should I learn Python or R for data science?
• How to move into AI, data science or DevOps
• StatQuest with Josh Starmer, OpenIntro Statistics and An Introduction to Statistical Learning, three resources the stages use.
• Related roadmaps: SQL and databases and MLOps and LLMOps.

What you can do afterwards

Apply for data scientist, analyst, analytics and research assistant roles; the roadmap is clear that these are real first steps, not consolation prizes. On the ML focus, you are ready to start the Machine learning roadmap with good validation habits.

FAQ

How long does it take to become a data scientist? The roadmap estimates about 215 hours of learning: roughly 24 weeks at 12 hours a week or 43 weeks at 5 hours a week, plus the job search.

Do I need strong maths for data science? You need to explain what your numbers mean rather than prove theorems. The statistics stage suggests explaining each concept out loud and simulating in code.

Python or R for data science? Python is the default; the Statistics and R focus exists for research-heavy work. The concepts are the same in both.

Is data science still worth learning in 2026? The roadmap leans on the parts that need judgement: cleaning, experiments, leakage checks and choosing thresholds from the cost of errors. Those are the conversations it says separate a data scientist from someone who ran a tutorial.