Data Scientist Career Guide: Experiments, Models, and Interviews

Two postings titled Data Scientist can describe jobs with little in common. One wants a SQL-fluent experimenter who can tell a product manager whether a new onboarding flow worked; the other wants someone to train, validate and ship a fraud model. Work out which one you are reading before you prepare, because the interviews differ too, though both lean hard on statistics.

Two jobs share one title

In analytics-heavy roles, often called product data science, you define what success means for a feature, design and read A/B tests, size opportunities, and answer causal questions from observational data when an experiment is impossible. In ML-heavy roles you build predictive models (credit risk at lenders, demand forecasts for retail and quick commerce, fraud detection, ranking), then work with engineers to put them into production and keep checking that they still work.

Here is how a week in a mixed role might run:

• A product manager asks whether a new referral reward "worked". It launched to everyone at once, so there is no control group; you estimate what you can and propose a staggered city-by-city rollout for the next change.
• You retrain a loan default model after a credit policy change, knowing that outcomes are only observed for approved applicants and the approved mix has shifted.
• You review an experiment readout that tracked fourteen metrics and celebrated the one that crossed p < 0.05.
• You present a one-page recommendation that gives a range, not a single number.

A data analyst explains what happened and keeps metrics trustworthy; a data scientist is expected to estimate what would happen under a change, and how confident anyone should be. An ML engineer makes models run reliably in production (serving, feature pipelines, retraining, monitoring), while the data scientist decides what to model, which metric proves it works, and whether it beats a simple rule. Read the verbs: "deploy, serve, Kubernetes, latency" points to the ML end; "experiment, stakeholders, metrics, insights" points to the analytics end.

What employers test beyond Python and machine learning

Most postings name Python, SQL and machine learning. What differs is how deeply each is examined.

Area	What good looks like	Weight
Statistics and experimentation	Power and sample size, confidence intervals, multiple comparisons, the peeking problem, sample ratio mismatch, variance reduction such as CUPED	Core everywhere; decisive in product roles
SQL	Cohort, funnel and retention queries written without help	Core; often the first screen
Python data stack	pandas, NumPy, scikit-learn and statsmodels, in analysis someone else can rerun	Core
Modelling	Linear and logistic regression you can interpret; gradient-boosted trees (XGBoost, LightGBM); validation that mirrors how the model will be used	Core in ML-heavy roles
Model evaluation	Precision and recall, PR-AUC for rare events, calibration, thresholds set from the cost of each error	Core for fraud, credit and churn
Communication	A recommendation with its uncertainty, on one page	Decides final rounds

Differentiators include causal inference (difference-in-differences, synthetic control, propensity methods), time series forecasting, Bayesian methods, and enough deployment knowledge (Docker, a small API, MLflow) to hand a model over cleanly. Deep learning matters for text, image and speech problems; on tabular business data, gradient-boosted trees remain the usual strong baseline. Some recent postings with this title also ask for LLM work, which overlaps with generative AI engineering; check how much of the role it is before reshaping your preparation.

How seniority is judged, and who hires data scientists

• 0–2 years: given a defined problem, you produce a correct analysis or a sound baseline model, write code others can follow, and say when the data cannot answer the question. Titles: Associate Data Scientist, Data Scientist I. Postings with this title accept freshers less often than analyst postings do, which is why many people start as analysts; the experience breakdown in the market snapshot on the India data scientist jobs page shows the current picture.
• 3–5 years: you frame problems in a domain, choose methods, design experiments, push back on questions the data cannot answer, and have models in production with engineering partners. Titles: Data Scientist II, Senior Data Scientist.
• 6+ years: you set methodology for others (experiment guidelines, metric definitions, model review standards), pick problems by business value, and mentor. Titles: Lead, Staff or Principal Data Scientist, Data Science Manager.

Services and analytics consulting firms put you on client problems across retail, banking and pharma, with more presentations and shorter projects; client work may include generative AI prototypes. Product companies and startups lean on experimentation and on models inside core flows such as pricing, ranking and fraud. GCCs of banks, insurers and retailers build risk, marketing and forecasting models for the parent company; at banks and insurers, model documentation and independent validation are part of the job.

The interview loop, round by round

1. Screen: which version of the job you have done, notice period, expectations.
2. SQL and Python: cohort and funnel queries and pandas manipulation, live or timed.
3. Statistics and probability: tests, distributions, conditional probability, experiment design.
4. Project walk-through: one past model or analysis, questioned on the target, the split, the metric, and what you would change.
5. Product or business case: define a metric, design an experiment, or answer a causal question without one.
6. Take-home or presentation: a dataset and a few days, then a 20–30 minute readout. The take-home assignments guide explains how to scope one.
7. Hiring manager: a stakeholder who wanted a different answer, and what you did.

Practice questions that reveal how you think

Each has a textbook answer and a better one. Panels listen for the better one: the assumption you check first, the trade-off you name, the action you would take.

1. "A checkout test shows p = 0.03 on day four of a planned 14-day run. The PM wants to ship today." Explain why stopping at a peek inflates false positives, raise weekly cycles and novelty effects, and offer a sequential test if early decisions genuinely matter.
2. "In a 50/50 split, control has 50,000 users and treatment 48,600." A chi-square test puts that imbalance far beyond chance. It is a sample ratio mismatch: find the assignment or logging bug before reading any metric.
3. "An experiment tracks 20 independent metrics at the 5% level, and nothing truly changed. How likely is at least one significant result?" About 64%, from one minus 0.95 to the power 20, which is why you pre-register a primary metric.
4. "Explain a confidence interval to a sales head in thirty seconds, without claiming a 95% chance that the true value lies inside it." A workable answer: our best estimate is a 4% lift, and the data are consistent with anything from 1% to 7%.
5. "Your churn model scores 0.92 AUC offline and does no better than a simple rule in production. Why?" Leakage from features recorded after churn, a random split on time-ordered data, a label that differs from the business definition, or training–serving skew.
6. "One transaction in 500 is fraud. A model catches 90% of fraud and wrongly flags 1% of genuine transactions. What share of alerts are real fraud?" Per 100,000 transactions, 180 true alerts sit among 998 false ones, so about 15%. Accuracy is useless here, since flagging nothing scores 99.8%; set the threshold from the precision–recall curve, the cost of each error, and how many alerts the review team can clear each day.
7. "When would you ship logistic regression over gradient boosting that scores slightly higher?" When decisions must be explained to customers or regulators, data is small, or stability matters more than a small lift.
8. "Write SQL for weekly retention by signup cohort." Then say whether week one means the calendar week or the seven days after each user's signup, and why it matters.
9. "A loyalty programme launched to every user in Hyderabad last month. Did it raise orders?" Without randomisation, compare against similar cities with difference-in-differences, check that pre-launch trends ran parallel, and account for festivals or a competitor's offer in the same weeks.

Resume bullets that show judgement

Reviewers read a data science resume for the decision your work informed, why the method fit, and evidence the impact was real, ideally from an experiment rather than an offline score. Put model metrics next to a business result, never instead of one. These are illustrative; the numbers stand in for your own.

• Example: "Designed and analysed 18 pricing and paywall experiments for a subscription app; introduced pre-registered primary metrics and power calculations, and identified a plan change that raised paid conversion by 6% (95% CI 3–9%)."
• Example: "Built a LightGBM default-risk model on 1.2 million loan applications with time-based validation and reason codes for declines; at the cut-off the credit team chose, back-testing showed 11% lower expected losses at the same approval rate."
• Example: "Replaced a manual weekly demand forecast for 300 dark stores with a hierarchical forecasting pipeline in Python, cutting forecast error from 31% to 19% WAPE; handed it to the ML platform team with drift monitoring."

Mistakes that cost interviews:

• A list of twenty algorithms, from KNN to GANs, with no problem attached to any of them.
• Titanic or Iris projects as the headline. Leave them off, or put them below real work.
• Accuracy reported on an imbalanced problem, or a 99% figure that invites the first leakage question.
• No sign of what changed in the business because of the work.

Routes into data science, and where it leads

From analyst: volunteer to analyse the next experiment, learn inference and regression properly rather than as library calls, and take one forecasting or segmentation problem end to end. From software or data engineering: your production skills are an advantage; the gap is usually statistics and evaluation design. If the systems side excites you more than the questions, the ML engineer path may fit better. Freshers: a degree in statistics, economics, mathematics or computer science helps, but projects that show judgement carry more weight than certificates.

Projects worth building:

• A peeking simulation: run a thousand simulated A/A tests, check significance daily, and count how often a "winner" appears by chance; then repeat with a fixed horizon. One notebook shows you understand why experiment rules exist.
• A forecasting backtest on public sales or energy-demand data, compared honestly against a seasonal naive baseline, with errors reported by segment.
• A decision-focused model: credit default on a public dataset, a threshold chosen from a cost matrix, a calibration plot, and a one-page memo for a non-technical reader.

From here, people move into staff-level data science, ML engineering, analytics leadership or product roles. Read the data engineer guide as well; knowing how your input tables are built prevents a whole class of silent errors.

Where to look for data science roles

Current openings: Bengaluru, Hyderabad, Mumbai, India nationwide and remote. A short routine for each application:

1. Sort the posting into analytics-heavy or ML-heavy by its verbs; How to Read a Software Job Description shows how to tell essentials from extras.
2. Move the matching project to the top of your resume, and check its layout against the ATS-friendly resume guide.
3. Rehearse two questions from the list above that match that version of the job.