A data scientist uses statistics and machine learning to answer questions with data, build predictive models and test whether changes work. The path is statistics and maths basics, Python with pandas, SQL, machine learning with scikit-learn, experiment design, then deploying one model. Analysts explain what happened, scientists predict and test, and ML engineers run models in production.
This page is for students and career switchers who want to work as a data scientist, and who want to know how the job differs from an analyst, an ML engineer and an AI engineer before they spend months on it. It gives the order of study, a free first resource for each skill, what to build, and hiring and pay evidence. Everything was opened and checked on 11 November 2026.
The topic order follows the AI and ML learning path. If you are not sure you need machine learning yet, start with the data analyst roadmap: SQL, statistics and storytelling are the base for both roles.
- 1Maths and statisticsprobability, inference
- 2Python and pandasload, clean, reshape
- 3SQLget data from databases
- 4Classical MLscikit-learn, validation
- 5Experiment designA/B tests, sample size
- 6Communicatelimits as well as results
- 7Deploy one modelan API or a demo
- 8Portfoliothree write-ups, one live
Data scientist, analyst, ML engineer, AI engineer: who does what
| Role | Main question | Typical output | Main tools |
|---|---|---|---|
| Data analyst | What happened, and why? | Queries, dashboards, recommendations | SQL, spreadsheets, BI tools |
| Data scientist | What will happen, and did the change work? | Models, experiments, analysis that informs a decision | Python, statistics, scikit-learn, SQL |
| ML engineer | How do we run this model reliably for users? | Services, pipelines, monitoring | Python, PyTorch, Docker. See the ML engineer roadmap |
| AI engineer | How do we build a product on an existing LLM? | Apps with retrieval, tools and evals | APIs, Python or TypeScript. See the AI engineer roadmap |
Titles overlap and change between companies, so read the posting, not the title. The AI vs data science vs machine learning page compares the fields, and Masai's AI and data science: how they work together is a longer read on how the fields relate.
Skills, why they matter, and the first free resource
| Skill | Why it matters | First free resource (checked 11 Nov 2026) |
|---|---|---|
| Maths: linear algebra, calculus, probability | Explains why models train, why they fail and what a loss function is doing | Mathematics for Machine Learning: the authors keep the PDF freely available |
| Statistics | Inference, uncertainty and bias separate a finding from noise. It is the core of the job | OpenIntro Statistics, free PDF, or Khan Academy Statistics and probability |
| Python and pandas | Almost all data science code is loading, cleaning and reshaping data | Kaggle Pandas: about 4 hours, no cost |
| SQL | Most data lives in databases, and you must pull your own training data | Kaggle Intro to SQL: about 3 hours, no cost |
| Machine learning with scikit-learn | Baselines, train and test splits, cross-validation and tuning come before any deep network | Kaggle Intro to Machine Learning, 3 hours, then the scikit-learn user guide and Google's Machine Learning Crash Course |
| Experiment design | Tells you if a product change worked, and how much data you need before you can say so | Udacity's A/B Testing course, listed as a free course taught by Google staff: five lessons, metrics, design and analysis |
| Communication | A model nobody trusts or understands is not used | storytelling with data blog |
| Deployment basics | A model that nobody can call is a notebook, not a product | FastAPI tutorial and the Hugging Face Spaces guide |
Experiment design: how big must a test be?
Data scientists are often asked whether a change moved a metric. Before running the test, you work out how many users you need, otherwise a real lift hides in noise. This script uses only Python's standard library to compute the visitors needed in each group of a two-sided test at the usual 5% significance and 80% power.
from statistics import NormalDist
def sample_size_per_group(p_base, p_new, alpha=0.05, power=0.8):
"""Visitors needed in each group of a two-sided A/B test on a conversion rate."""
z = NormalDist().inv_cdf
z_alpha = z(1 - alpha / 2)
z_beta = z(power)
variance = p_base * (1 - p_base) + p_new * (1 - p_new)
n = (z_alpha + z_beta) ** 2 * variance / (p_new - p_base) ** 2
return int(n) + 1 # round up
# Baseline sign-up rate 10%. Is a rise to 11% detectable?
print(sample_size_per_group(0.10, 0.11))
# A one-point lift on a 2% baseline needs a much larger test.
print(sample_size_per_group(0.02, 0.03))
# Halve the effect and the sample size roughly quadruples.
print(sample_size_per_group(0.10, 0.105))It prints 14,749, then 3,823, then 57,760. A one-point lift from 10% to 11% needs about 14,700 visitors per group, and halving the lift roughly quadruples the requirement. Calculators use slightly different formulas, so expect small differences. The lesson matters more than the number: small effects need big samples, so decide the minimum effect worth detecting before you start.
What to build
- A tabular prediction with an honest baseline. Customer churn or house prices. Report a simple baseline first, then a better model, with cross-validation and one metric chosen for the business, not just accuracy. The AI and ML project ideas for beginners page lists datasets and metrics.
- An experiment analysis. Simulate or download A/B test data, compute the sample size, run the test and write a one-page recommendation that says what you would and would not conclude.
- One deployed model. Wrap a trained model in a FastAPI service on Render or a Gradio demo on Hugging Face Spaces. A live link is worth more than a notebook.
Find data in the free datasets list. The Kaggle Intro to Machine Learning course ends with a lesson on competitions, a safe place to compare your approach with others. Masai's 25 data science projects for students is a longer list.
2027 outlook, entry level and pay
US projections, for scale. The US Bureau of Labor Statistics projects data scientist employment to grow 35% from 2025 to 2035, from 275,600 jobs in 2025, with about 24,800 openings a year. It lists a bachelor's degree as the typical entry requirement and notes some employers prefer a master's or doctorate. That is US data, not India.
India, September 2026. Naukri's JobSpeak shows AI and ML hiring up 20% year on year against 2% for all white-collar hiring. Within AI and ML, the 13 to 16 year band grew 54% and the 8 to 12 year band 28%, while fresher AI and ML hiring grew 19%. These are growth rates, not counts. Our reading: demand is rising for everyone but fastest for experienced people, so a fresher should expect competition and aim for roles that combine analysis, SQL and modelling.
Entry level and AI. The Stanford Digital Economy Lab paper, revised on 12 August 2026, finds US employment of 22 to 25 year olds in AI-exposed occupations 19% below trend relative to less-exposed peers, with no comparable gap for experienced workers. The authors describe it as early and descriptive. Do not assume it applies to India, but do build skills AI tools do not replace: framing the question, testing, and explaining limits.
Pay. On levels.fyi, self-reported data scientist compensation in India has a median total of about ₹30.0 lakh from 1,023 salaries, with the 25th percentile at ₹19.1 lakh, the 75th at ₹44.1 lakh and the 90th at ₹61.6 lakh (page updated 11 November 2026). It covers all levels and people who chose to post, so it is not a fresher figure.
Interview preparation
- Statistics. Explain a p-value and a confidence interval in plain words, Type I and Type II errors, and why you cannot peek at an A/B test every day.
- Machine learning. Overfitting and how to detect it, data leakage, how to choose a metric. A classic: a fraud model scores 98% accuracy, is that good? Not if 98% of transactions are genuine, because predicting 'genuine' for everything scores the same. Ask about class balance, then discuss precision and recall.
- SQL and pandas. Joins, window functions, group by, and cleaning a messy file.
- A case or take-home. Frame the question, pick a metric, build a baseline, state limits. Many companies give a dataset and a few days.
- Your projects. Be ready to explain every modelling choice, including why you did not use something more complex.
Frequently asked questions
What is the roadmap to become a data scientist?
Learn statistics and basic maths, Python with pandas and SQL, then machine learning with scikit-learn and experiment design. Build three projects, one of them deployed, and write each up. Add deep learning only if the roles you target ask for it.
Do I need a master's degree to become a data scientist?
Not always. The US Bureau of Labor Statistics lists a bachelor's degree as the typical entry requirement and says some employers require or prefer a master's or doctorate. Postings in India vary, so check the roles you want and let projects and statistics knowledge carry weight.
How much maths do I need for data science?
Enough probability and statistics to reason about uncertainty, plus linear algebra and basic calculus to understand how models learn. You do not need advanced proofs for most applied roles. Statistics matters first, then the rest as you move into machine learning.
Can a fresher become a data scientist?
Yes, but entry roles are competitive, and data scientist postings often ask for experience. Many freshers start as data analysts or in analytics-heavy roles and move towards modelling over time. Show finished projects with baselines, metrics, a written conclusion and a deployed demo.
What is the difference between a data scientist and a machine learning engineer?
A data scientist frames questions, analyses data, builds models and runs experiments. An ML engineer takes models into production and keeps them reliable with pipelines, serving and monitoring. At small companies one person may do both jobs, so ask what a posting holds you responsible for.
Is data science still a good career in 2027?
Demand data is mixed by level. US projections are strong, Naukri shows AI and ML hiring rising faster than the overall market, and early-career hiring in AI-exposed jobs looks weaker in US payroll data. Build skills that combine statistics, SQL and modelling, not just tool knowledge.
Sources checked
- BLS data scientist outlook 2025 to 2035
- Self-reported India data scientist pay
- Naukri JobSpeak September 2026
- Intro to Machine Learning: 3 hours, no cost
- Free A/B testing course by Google staff
- Stanford paper revised 12 Aug 2026
Learn this with mentors
AI and machine learning programmes with admissions open: live classes, mentors and project work.

