On MyInternships.in, NumPy is listed by 8,496 candidates and Pandas by 8,056. Machine Learning is listed by 3,281. Divide the NumPy figure by the Machine Learning one and you get about 2.6.
In other words, a large group of students on our platform have the data-handling tools and have not claimed the machine learning label. Whether you treat that group as underqualified or as your best source of ML interns will shape your whole search.
The ML stack, as listed by candidates
All numbers come from our registered candidates' own skill entries, counted on 1 October 2026. Someone listing four of these skills is in four rows, and a listed skill is a claim that still needs checking.
| Skill listed on profile | All India | Bengaluru | Pune | Chennai | Hyderabad | Mumbai |
|---|---|---|---|---|---|---|
| Python | 21,874 | 2,458 | 1,778 | 1,550 | 1,968 | 899 |
| NumPy | 8,496 | 914 | 782 | 565 | 767 | 362 |
| Pandas | 8,056 | 861 | 736 | 548 | 751 | 359 |
| Machine Learning | 3,281 | 344 | 229 | 208 | 199 | 152 |
| Deep Learning | 975 | 128 | 73 | 75 | 77 | 39 |
| NLP | 864 | 117 | 67 | 68 | 71 | 44 |
The stack narrows as it becomes more specialised. Python is on 21,874 profiles. Machine Learning, at 3,281, is 15 per cent of that. Deep Learning is on 975 and NLP on 864.
Bengaluru leads Machine Learning with 344 listings and Pune has 229. One detail is worth noting: Chennai, with 208, is ahead of Hyderabad at 199, even though Hyderabad has more Python listings overall (1,968 against 1,550).
What an ML intern really does all day
Students imagine training neural networks. In most companies, an intern on a machine learning team spends the bulk of the time on less glamorous work:
- pulling and joining data from several sources;
- cleaning it and discovering why it is wrong;
- building simple baselines;
- running evaluations and writing up what happened;
- labelling or reviewing examples.
That is why the Pandas and NumPy pool matters. A careful student who is excellent at data handling and knows basic statistics is frequently more useful than one who can recite architectures and has never questioned a dataset.
Put this description in your posting. It discourages candidates who want only the fashionable part, and it reassures capable students who assumed they were not advanced enough.
Separating understanding from imitation
Machine learning projects on fresher profiles tend to look alike, since many originate from a small set of public notebooks. To find understanding, ask about decisions.
- "Why did you choose that metric?" If the classes were imbalanced and they reported plain accuracy, do they see the issue once you point at it?
- "What was your baseline?" A candidate who compared a model against a trivial rule understands evaluation.
- "How did you split the data, and why that way?"
- "What did the model get wrong?" Anyone who looked at individual errors has done real work.
Rehearsed definitions are not what you are after. A candidate who says "I did not think about that at the time, but now I would" is showing the capacity to learn, which is the point of an internship.
The practical test: a small table with a leak
Create an invented dataset of a few thousand rows about, say, whether a customer renews a subscription. Include about ten ordinary columns, some missing values, and one column that leaks the answer: for example, a "cancellation call logged" flag that is only ever filled in after the customer has already left.
The brief:
Predict renewal. Deliver a notebook and a one-page summary for a non-technical manager covering what you built, how well it works, which factors matter, and what you would be cautious about. A simple model explained well is preferred to a complex one explained badly.
Allow a weekend, with an expected effort of three hours.
What the leak tells you
Submissions fall into three groups.
- Near-perfect score, no questions asked. The candidate used the leaking column and celebrated. This is the commonest outcome and the most instructive, because unquestioned perfect results are exactly how flawed models reach production.
- Near-perfect score, with suspicion. "The score is unusually high; the cancellation flag seems to dominate and may not be available at prediction time." This candidate has the right reflex even if they left the column in.
- Column removed with an explanation. The strongest result. The reported score will be lower and more honest.
Do not treat the first group as automatic rejections. In the follow-up call, show them the feature importance and ask what they think of it. Some will work it out in a minute, which is a perfectly good sign in a beginner.
Beyond the leak, check the basics: a held-out test set, a baseline, sensible handling of missing values, and a metric suited to the problem. Then read the one-page summary as the manager would. Does it avoid jargon, does it state limits, and would you know what to do next after reading it?
Compute, data access and expectations
Decide in advance what data an intern may see, and prepare an anonymised or sampled set if the real one is sensitive. Agree what computing resources they can use and who approves anything costly.
Define success as a clear, reproducible analysis, not a headline accuracy. Interns pushed to hit a number learn to overfit; interns asked to explain a result learn to think.
Compare ML supply across cities with recruiter intelligence and see employer tools for posting templates. When your dataset is ready, hire machine learning talent on MyInternships.in, and consider inviting strong Pandas candidates who never used the ML label.


