A realistic learning path into data science, from someone doing the job

If you are studying part time, around ten hours a week, budget ten to twelve months to be genuinely employable. The order I would follow is Python, SQL, pandas, statistics, machine learning, then one project you actually deploy. The order matters more than which course you pick, and the two stages people skip, SQL and deployment, are the two that decide interviews.

I am Bablu Kumar Jha. I run BeeProton AI and work as a Data Scientist at Brijwasi Sweets (India) Pvt Ltd. I mention the day job only so you know where this comes from: most of my week is spent pulling data, arguing with it about whether it is correct, and explaining a number to someone who has to make a decision with it. Model training is a small slice. That gap between the syllabus and the work is what this article is about.

The sequence

Months	Focus	Done when you can
1–2	Python	Write a 150-line script that reads files, handles errors and does not need rewriting
3	SQL	Answer a business question with a join, a GROUP BY and a window function
4	pandas and cleaning	Take a messy CSV to an analysis-ready frame and explain every decision
5	Statistics	Say what a p-value means, and why a 2 percent lift on 80 visitors means nothing
6–7	Machine learning	Build a model with honest validation and explain why it beats the obvious baseline
8	One deployed project	Send someone a URL where your work runs
9–10	Second project, interview prep	Do live SQL and pandas under mild pressure

Treat the right-hand column as the test. Finishing a course is not the same as passing it. Our learning tracks follow roughly this order, and the roadmaps lay out the topic lists if you would rather see them all at once.

Months 1 to 2, Python

Lists, dicts, functions, files, exceptions, comprehensions, and enough object-oriented code to read other people's. Skip decorators, metaclasses and async for now; they will not appear in an analytics interview.

Build something with input you did not create: parse your bank statement CSV and total by month. Real data has blank rows and inconsistent dates, and handling that in month one saves you months later.

Month 3, SQL, earlier than you think

Put SQL third, not last, for the reasons in the section below.

Month 4, pandas

Selection with .loc, groupby, merge, reshaping with pivot_table and melt, dt accessors for dates. The skill that gets you hired is not knowing the methods, it is knowing what to do when a merge that should produce 4,000 rows produces 4,300.

before = len(orders)
out = orders.merge(customers, on="customer_id", how="left", validate="m:1")
assert len(out) == before, f"merge changed row count: {before} -> {len(out)}"

validate="m:1" makes pandas raise when the right-hand key is not unique, catching duplicated customer rows at the point of the mistake instead of in a total that is quietly 8 percent too high. That habit is what separates a working analyst from someone who has finished a pandas course.

Month 5, statistics

Descriptive statistics, distributions, sampling and standard error, confidence intervals, hypothesis tests, correlation against causation. Do it with data rather than proofs: analyse an A/B test on a public dataset and write the conclusion in five sentences a manager could act on.

Months 6 to 7, machine learning

Linear and logistic regression, decision trees, random forests, gradient boosting, k-means. For each: what it assumes, what it needs from the data, how it fails. Then evaluation, which carries more weight than the algorithms. Train and test splits that respect time order, cross-validation, the difference between accuracy and recall when 3 percent of rows are the positive class, and always a dumb baseline to beat. A churn model at 94 percent accuracy is worthless if 94 percent of customers do not churn.

Month 8, deploy one thing

A Streamlit or FastAPI app on a free tier, or a Django page if you want the web fundamentals too. It does not have to be impressive. It has to be reachable by a URL, handle a bad input without a traceback, and still work in three months. This step moves you past most other applicants, because notebooks are common and running software is not.

How much maths you actually need

Less than the course catalogues imply, more than zero.

Needed properly: mean against median and when each lies, variance and standard deviation, probability basics including conditional probability, the normal distribution and the central limit theorem in outline, sampling error, and the logic of a hypothesis test. These come up almost weekly.

Useful to understand, not to derive: what a vector and a matrix multiplication are, why a dot product measures similarity, what a gradient is and why following it downhill fits a model. You want enough intuition that "the loss is not decreasing" means something to you.

Optional for most jobs: proving convergence, deriving backpropagation by hand, measure-theoretic probability. Research roles and some product-company interviews do ask, and if that is your target, add linear algebra and calculus properly. For an analytics or applied data science role in India, the maths bar sits below the SQL bar and the communication bar.

Why SQL decides Indian analytics interviews

Because the data lives in a database and the job is mostly getting it out. In practice most analytics teams here run MySQL, PostgreSQL or a warehouse like BigQuery or Snowflake, and the daily work is a query, not a notebook. When someone asks how sales moved week on week across four branches, the answer is SQL, and pulling everything into pandas to do it marks you out immediately.

It is also the easiest thing to test live, so interviewers test it. Expect a window function:

SELECT branch,
       week,
       revenue,
       revenue - LAG(revenue) OVER (PARTITION BY branch ORDER BY week) AS wow_change
FROM weekly_sales
ORDER BY branch, week;

If LAG, ROW_NUMBER, RANK and a PARTITION BY are not comfortable, that is your highest-return week of study. Add LEFT JOIN against INNER JOIN and what happens to NULLs, HAVING against WHERE, and why COUNT(*) and COUNT(column) differ. Practise on our SQL track and then solve problems under time pressure on Practice, because knowing the syntax and producing it in eight minutes while someone watches are separate skills.

Portfolio projects that get replies

Three characteristics, and most projects have none of them.

Messy input you did not clean for yourself. The Titanic dataset tells a recruiter nothing: it is already clean and 200,000 people have done it. Pull something raw from the government open data portal, or a set of daily NAV files, and show the cleaning.

A decision at the end. Not "I achieved 0.89 AUC" but "following the model's top 200 flagged accounts would catch roughly three quarters of the churn, and here is what it costs to be wrong." A number somebody can act on.

Deployed, with a README that starts with the finding. Three sentences on what you found, then a screenshot, then how to run it. Most reviewers spend 90 seconds on a repository.

Four that work well in an Indian context: rainfall against crop yield by district, one commodity's mandi prices with a simple forecast and an honest error figure, a GST invoice extractor turning PDFs into a reconciliation sheet, and a bus or train punctuality dashboard from published timetable data. Each has real cleaning work and no chance of being the fiftieth identical submission.

Two finished projects beat eight abandoned ones. A repository of half-done notebooks reads as a habit of not finishing.

What interviews ask

Four rounds, typically.

1. Screening. Walk through a project in five minutes. Rehearse it. Most candidates ramble, and the ones who state the question, the data, the approach and the result in that order stand out immediately.
2. SQL, live. Two or three problems, usually a join plus an aggregation plus a window function. Talk while you type.
3. Python and pandas. Clean a small messy frame, or write a function against a spec. Rarely hard data structures for analytics roles.
4. Statistics and ML reasoning. Why this model, how did you validate it, what does this p-value mean, how would you handle a class imbalance, what if the client says the numbers look wrong.

The question that sinks people is the follow-up: "why that choice, and what would you do differently?" It exists to find out whether you understood your own project or followed a tutorial. Prepare one honest answer about something you got wrong and fixed.

The honest timeline

At 8 to 10 hours a week, ten to twelve months to being a credible junior candidate. At 20 hours a week, six to seven. Full time with real focus, four to five, though that shortens learning and not experience, and it shows when an interview goes one layer deeper than the course did.

Two points where people stop. Month 4, when statistics feels like a detour from the machine learning they came for. Month 8, when the model works in the notebook and deployment turns out to involve environments, dependency versions and a build log. Both are worth pushing through, and the second is where most of your advantage over other applicants gets built.

A job search here can take three to six months on top of the learning, with long silences that are not about you. Keep building something small through it, because "what have you worked on recently" is easier to answer when the answer is this month.

Frequently asked questions

Do I need a degree in statistics or computer science? No, though it shortens the screening stage. Without one you need a portfolio that proves the skills and you should expect more first-round rejections. People move in from mechanical engineering, commerce and teaching every year.

Are certificates worth anything? As a syllabus and a deadline, yes. As a hiring signal, very little on their own. Nobody in an interview has ever asked me about a certificate; they ask about the project.

Should I learn R or Python? Python, for the Indian job market. R remains excellent for statistics and is still used in research and some pharma work, but Python is what postings ask for and what fits into production systems.

Where does AI and LLM work fit in? After this sequence, not instead of it. Retrieval systems and prompt-driven pipelines still stand on data handling, evaluation and judgement about whether an output is correct. Those are exactly the skills in months 3 to 7.