What Interviewers Actually Look for in a Fresher Data Scientist

A FRESHER PROFILE INTERVIEWERS TEND TO TRUST

Skills (honest, short list):
  Python (pandas, numpy, scikit-learn), SQL,
  statistics, one visualisation tool

Projects (two, not ten):
  1. End-to-end supervised ML project on real, messy data
     with a baseline, error analysis and a short write-up
  2. An analysis project that answers a business question
     with SQL and clear charts

Evidence of learning:
  A README that explains a problem you got stuck on
  and how you solved it

EXAMPLE FRESHER TAKE-HOME (format varies by company)

You are given a CSV of past loan applications.

1. Explore the data and describe any quality issues.
2. Build a model to predict default.
3. Explain which metric you chose and why.
4. Summarise your findings for a non-technical manager
   in under one page.

What reviewers usually look at:
- Did you check missing values and odd entries first?
- Is there a baseline?
- Is the metric suited to an imbalanced target?
- Is the summary readable without code?

import pandas as pd

orders["order_date"] = pd.to_datetime(orders["order_date"])

monthly = (orders
           .assign(month=orders["order_date"].dt.to_period("M"))
           .groupby(["month", "city"], as_index=False)
           .agg(orders=("order_id", "nunique"),
                revenue=("amount", "sum")))

top3 = (monthly.sort_values(["month", "revenue"], ascending=[True, False])
               .groupby("month")
               .head(3))

before = len(orders)
enriched = orders.merge(customers, on="customer_id",
                        how="left", validate="many_to_one")
assert len(enriched) == before, "merge duplicated rows"

by_mode = orders.pivot_table(index="city", columns="payment_mode",
                             values="amount", aggfunc="sum",
                             fill_value=0)

-- Monthly revenue by city (PostgreSQL)
SELECT DATE_TRUNC('month', order_date) AS month,
       city,
       COUNT(DISTINCT order_id) AS orders,
       SUM(amount) AS revenue
FROM orders
GROUP BY 1, 2;

-- Top 3 cities per month
WITH monthly AS (
    SELECT DATE_TRUNC('month', order_date) AS month,
           city, SUM(amount) AS revenue
    FROM orders
    GROUP BY 1, 2
), ranked AS (
    SELECT *, ROW_NUMBER() OVER (PARTITION BY month
                                 ORDER BY revenue DESC) AS rn
    FROM monthly
)
SELECT month, city, revenue FROM ranked WHERE rn <= 3;

-- Payment mode breakdown
SELECT city,
       SUM(CASE WHEN payment_mode = 'UPI'  THEN amount ELSE 0 END) AS upi,
       SUM(CASE WHEN payment_mode = 'CARD' THEN amount ELSE 0 END) AS card
FROM orders
GROUP BY city;

import pandas as pd

# The bug: a warning, and the new column sometimes missing
high = df[df["amount"] > 1000]
high["flag"] = 1          # SettingWithCopyWarning

# What I learned: filtering can return a view or a copy,
# so assigning to it may not behave as expected.

# Fix 1: take an explicit copy when I want a new DataFrame
high = df[df["amount"] > 1000].copy()
high["flag"] = 1

# Fix 2: assign on the original with .loc when I want to
# change df itself
df.loc[df["amount"] > 1000, "flag"] = 1
df["flag"] = df["flag"].fillna(0).astype(int)

Interviewer: Have you used target encoding?

Weak: No.

Strong: "I have not used it directly. From the name, I
would guess it replaces a category with a statistic of
the target, like the average default rate for each city.

If so, my first worry would be leakage: if I compute the
average using the same rows I train on, the model sees
its own labels. I would compute it inside cross-validation
folds.

Is that close to how your team uses it?"

(Reasoning from first principles, plus a sensible risk,
 is exactly the learning signal interviewers want.)

CAN YOU EXPLAIN, IN UNDER A MINUTE EACH?

[ ] Mean, median and when each misleads
[ ] Standard deviation and why spread matters
[ ] Hypothesis test, p-value, confidence interval
[ ] Train, validation and test sets
[ ] Overfitting and how to detect it
[ ] Precision, recall and when each matters more
[ ] Linear vs logistic regression
[ ] How a decision tree and a random forest differ

CAN YOU DO THESE WITHOUT SEARCHING?

[ ] Read a CSV, inspect shape, dtypes, nulls
[ ] Parse dates and extract month or weekday
[ ] groupby with several aggregations
[ ] merge and verify the row count
[ ] pivot_table with fill_value
[ ] SQL GROUP BY with HAVING
[ ] SQL LEFT JOIN and spot missing matches
[ ] ROW_NUMBER for top-N per group
[ ] Train/test split and a baseline model

FOR YOUR MAIN PROJECT

[ ] Real data with real problems (not a perfectly clean set)
[ ] Data quality issues described and handled
[ ] Baseline model compared with final model
[ ] Metric chosen for a stated reason
[ ] Error analysis: where does the model fail?
[ ] README a non-technical person can follow
[ ] You can explain every line of the code
[ ] One thing you would do differently next time