Data Scientist Career Path in India: Skills, Roadmap & Salary 2026

"Data scientist" covers a surprisingly wide range of jobs in India. At one company it means building credit risk models for a lending app. At another it means running A/B tests on a checkout flow. At a third it is a dashboard-heavy analyst role with a more glamorous title. Before planning a career path, it helps to know which of these you are aiming for, because the skills, interviews and salaries differ.

This guide covers the role as it exists at Indian product companies, fintechs, e-commerce firms, GCCs and consulting firms: what the work is, how it compares with adjacent roles, a skills roadmap with code, how to build a portfolio, the interview loop, approximate salary bands and the realistic entry routes.

What Data Scientists Actually Do

A typical week for a data scientist at, say, a food delivery or fintech company might include:

• Writing SQL to pull and clean data from the warehouse (often the biggest time sink).
• Designing and analysing an A/B test on a new feature or pricing change.
• Building or improving a model: delivery time prediction, fraud scoring, churn propensity, recommendation ranking, demand forecasting.
• Investigating a metric drop ("orders fell 6% in Pune last week, why?").
• Presenting findings to product managers and business leaders, and recommending a decision.
• Working with ML or data engineers to put a model into production and monitor it.

The common thread is decision-making under uncertainty. The most valued data scientists are not the ones with the fanciest models; they are the ones whose analysis changes what the company does, and who can say how confident they are.

Data Scientist vs Data Analyst vs ML Engineer

Two more roles sit nearby: analytics engineers (modelling data in the warehouse with SQL and tools like dbt) and data engineers (building pipelines and platforms; see our data engineer career path). If you enjoy building systems more than analysing results, compare the ML engineer path in our AI/ML engineer vs SDE guide.

The Skills Roadmap, in Order

Stage 1: SQL is the most tested skill

Almost every data science interview in India includes SQL, and many candidates underprepare for it. You should be fluent with window functions, self-joins, date logic and cohort queries. Here is a typical interview-style question: for each month, what share of new users placed a second order within 30 days of their first?

WITH first_orders AS (
  SELECT user_id, MIN(order_date) AS first_date
  FROM orders
  GROUP BY user_id
),
repeat_flag AS (
  SELECT f.user_id,
         DATE_TRUNC('month', f.first_date) AS cohort_month,
         EXISTS (
           SELECT 1
           FROM orders o
           WHERE o.user_id = f.user_id
             AND o.order_date > f.first_date
             AND o.order_date <= f.first_date + INTERVAL '30 days'
         ) AS repeated
  FROM first_orders f
)
SELECT cohort_month,
       COUNT(*) AS new_users,
       ROUND(100.0 * AVG(repeated::int), 1) AS repeat_rate_pct
FROM repeat_flag
GROUP BY cohort_month
ORDER BY cohort_month;

Interviewers look for whether you define the cohort correctly (by first order, not any order), handle the 30-day window precisely, and mention caveats: the most recent cohort has not had a full 30 days yet, so its rate is biased downward. That last point is the kind of thinking that separates a data scientist from a query writer. Practise more in our SQL interview questions.

Python for data work

Know pandas well (groupby, merge, pivot, time series resampling, handling missing values), NumPy basics, and one plotting library. Write functions, not 400-cell notebooks. Our Python for ML guide covers the essentials.

Statistics and A/B Testing: The Core of the Job

Statistics is where many self-taught candidates are weakest and where interviews probe hardest. You need probability (conditional probability, Bayes' rule, common distributions), sampling and the central limit theorem, confidence intervals, hypothesis testing and p-values (including what a p-value is not), power and sample size, and linear and logistic regression with interpretation. For the maths behind this, see math for AI and ML.

Worked example: reading an A/B test

A food delivery app tests a new checkout button. Control: 10,000 users, 1,200 conversions (12.0%). Treatment: 10,000 users, 1,290 conversions (12.9%). Is the difference significant?

1. Pooled conversion rate: p = (1,200 + 1,290) / 20,000 = 0.1245.
2. Standard error of the difference: SE = √(p(1 − p)(1/10,000 + 1/10,000)) = √(0.1245 × 0.8755 × 0.0002) ≈ 0.00467.
3. z-statistic: (0.129 − 0.120) / 0.00467 ≈ 1.93.
4. Two-sided p-value ≈ 0.054, just above the conventional 0.05 threshold.

The naive conclusion is "not significant, the feature failed." A strong data scientist says more: the observed lift is 0.9 percentage points (7.5% relative), the 95% confidence interval for the difference is roughly −0.02 to +1.82 points, so the test is underpowered to detect an effect of this size rather than evidence of no effect. The right questions next are: what was the pre-registered minimum detectable effect, was the sample size planned for it, and is it worth extending the test? Also check guardrail metrics (order value, cancellations, latency) and sample ratio mismatch.

Experimentation pitfalls interviewers ask about

• Peeking: repeatedly checking and stopping when p < 0.05 inflates false positives.
• Multiple comparisons: testing 20 metrics and celebrating the one that moved.
• Novelty effects and day-of-week seasonality: run tests for full weeks.
• Network effects: in marketplaces (riders and restaurants, drivers and passengers), user-level randomization can be biased; consider city- or time-based switchback designs.
• Simpson's paradox when segments (say, metro vs tier-2 cities) are mixed in different proportions.

Machine Learning Skills That Matter in Industry

Most industry data science problems in India are tabular: transactions, users, orders, events. Gradient-boosted trees and well-engineered features solve a large share of them. Deep learning matters for text, images, recommendations and LLM-based products.

from sklearn.compose import ColumnTransformer
from sklearn.impute import SimpleImputer
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import roc_auc_score
from sklearn.model_selection import train_test_split
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import OneHotEncoder, StandardScaler

numeric = ["tenure_months", "monthly_spend"]
categorical = ["city_tier", "plan"]
X, y = df[numeric + categorical], df["churned"]

# Split FIRST, so imputation and scaling statistics come only from training data
X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.2, stratify=y, random_state=42)

preprocess = ColumnTransformer([
    ("num", Pipeline([("impute", SimpleImputer(strategy="median")),
                      ("scale", StandardScaler())]), numeric),
    ("cat", OneHotEncoder(handle_unknown="ignore"), categorical),
])
model = Pipeline([("prep", preprocess),
                  ("clf", LogisticRegression(max_iter=1000, class_weight="balanced"))])

model.fit(X_train, y_train)
probs = model.predict_proba(X_test)[:, 1]
print(f"Test ROC-AUC: {roc_auc_score(y_test, probs):.3f}")

Why this matters: fitting a scaler or imputer on the full dataset before splitting leaks test-set information into training and inflates your metrics. Wrapping preprocessing in a Pipeline guarantees the same transformations are learned on training folds only, including during cross-validation. For churn with a time dimension, you would also split by time rather than randomly. For a deeper treatment of models, see classical machine learning.

Building a Portfolio That Gets Interviews

Recruiters have seen the Titanic survival notebook and the Iris classifier thousands of times. Projects that stand out look like work: a business question, messy data, a defensible method and a recommendation.

What each project should include

• A one-paragraph problem statement framed as a business decision.
• Data cleaning decisions and their justification.
• A simple baseline before any complex model.
• Metrics tied to the decision, not just accuracy.
• Limitations and what you would do with more time or data.
• A short written summary a non-technical manager could read in two minutes.

Kaggle competitions are useful for learning modelling techniques, and a good ranking helps. But a portfolio of only competitions can look narrow, because real work involves defining the problem and the metric, which competitions do for you.

The Data Science Interview Process

Processes vary by company; confirm with your recruiter. A common loop at Indian product companies and GCCs:

How to answer a metric investigation case

Take "daily orders dropped 10% yesterday." A structured answer:

1. Clarify: how was it measured, compared to what baseline (yesterday vs same day last week), is it all platforms and cities?
2. Rule out data issues: logging changes, pipeline delays, a broken event after an app release.
3. Check external factors: holidays, weather, a cricket match, a competitor's sale, payment gateway outages.
4. Decompose the metric: orders = sessions × conversion; segment by city, platform, app version, new vs returning users, payment method.
5. Form hypotheses and test them against the segments that moved most.
6. Recommend an action and how you would confirm the fix.

For a broader set of ML interview questions, see our AI and ML interview prep guide.

Data Scientist Salary Bands in India (2026)

These are approximate 2026 ranges from public sources such as AmbitionBox, Glassdoor and levels.fyi; they vary substantially by company, team, city and negotiation, and total compensation at product companies may include bonuses and stock.

Data analyst salaries typically start lower than data scientist salaries, and ML engineer salaries at product companies are often at or above data scientist levels for equivalent experience. Bengaluru, Hyderabad, Pune, Gurugram and Mumbai concentrate the highest-paying roles. For negotiation tactics, see our salary negotiation guide.

Realistic Ways Into Data Science

Be cautious with expensive short courses that promise a data scientist job. Most hiring managers care about what you can do in the SQL, statistics and case rounds, and what your portfolio shows, not certificates.

How GenAI Is Changing the Data Scientist Role

Large language models are changing day-to-day data science work in India and globally. Text classification, extraction and summarization tasks that once required a custom NLP model can often start with an LLM baseline. Code assistants speed up SQL and pandas work. At the same time, new work has appeared: evaluating LLM outputs, designing experiments for AI features, building retrieval systems and measuring their quality.

The skills that grow in value are the ones LLMs do not replace: problem framing, experimental design, causal reasoning, evaluation methodology and judgment about data quality. If you want to add GenAI skills, start with RAG and LLM evaluation; evaluation in particular is a natural extension of a data scientist's statistical toolkit.

Frequently Asked Questions

Pranjal Jain

Ex-Microsoft SDE · IIT Kanpur · Founder of Prepflix. Helps engineers crack startup and product company interviews across India.