What Interviewers Actually Look for in a Fresher ML Engineer

Q: What is overfitting?

A: Overfitting is when the model learns the training data
too well, including noise, and does not generalise to new
data. We can reduce it with regularisation, dropout or more
data.

(Correct, but it is a textbook line. I still do not know
whether this person has ever seen overfitting happen.)

Q: What is overfitting?

A: In my loan-default project, a decision tree with no depth
limit scored almost perfectly on training data but much lower
on validation. That gap was overfitting: the tree memorised
individual applicants. Limiting max_depth and setting
min_samples_leaf closed most of the gap, and validation
score went up even though training score went down.

from sklearn.datasets import make_classification
from sklearn.model_selection import train_test_split
from sklearn.tree import DecisionTreeClassifier

X, y = make_classification(n_samples=2000, n_features=20,
                           n_informative=5, flip_y=0.1,
                           random_state=0)
X_tr, X_va, y_tr, y_va = train_test_split(X, y, random_state=0)

for depth in [2, 4, 6, 10, None]:
    tree = DecisionTreeClassifier(max_depth=depth, random_state=0)
    tree.fit(X_tr, y_tr)
    print("depth=%-4s train=%.3f valid=%.3f" % (
        depth, tree.score(X_tr, y_tr), tree.score(X_va, y_va)))

import pandas as pd

df = pd.DataFrame({
    "city": ["Noida", "noida ", "Delhi", "Delhi", None, "Pune"],
    "amount": ["1,200", "800", "450", None, "300", "2,000"],
    "defaulted": [0, 1, 0, 1, 0, 0],
})

# 1. Inspect before changing anything
print(df.dtypes)
print(df.isna().sum())

# 2. Clean text and numbers
df["city"] = df["city"].str.strip().str.title().fillna("Unknown")
df["amount"] = pd.to_numeric(df["amount"].str.replace(",", ""),
                             errors="coerce")

# 3. Group-wise default rate and median amount
print(df.groupby("city").agg(
    default_rate=("defaulted", "mean"),
    median_amount=("amount", "median"),
    rows=("defaulted", "size")))

import numpy as np
import pandas as pd

def add_amount_features(df: pd.DataFrame) -> pd.DataFrame:
    """Add clean amount and a log version. Does not modify input."""
    out = df.copy()
    out["amount"] = pd.to_numeric(
        out["amount"].astype(str).str.replace(",", ""), errors="coerce")
    out["amount"] = out["amount"].fillna(out["amount"].median())
    out["log_amount"] = np.log1p(out["amount"])
    return out

# Run with: pytest -q
import math
import pandas as pd
from features import add_amount_features

def test_commas_and_missing_are_handled():
    df = pd.DataFrame({"amount": ["1,000", None, "3,000"]})
    out = add_amount_features(df)
    assert out["amount"].tolist() == [1000.0, 2000.0, 3000.0]
    assert not out["log_amount"].isna().any()

def test_input_is_not_modified():
    df = pd.DataFrame({"amount": ["5"]})
    add_amount_features(df)
    assert df["amount"].tolist() == ["5"]

def test_log_value():
    out = add_amount_features(pd.DataFrame({"amount": ["0"]}))
    assert math.isclose(out["log_amount"][0], 0.0)

WEEKS 1-2: FUNDAMENTALS WITH OWN EXAMPLES
- Overfitting, underfitting, bias and variance
- Leakage, class imbalance, scaling
- For each: one line definition + one example you ran

WEEKS 3-4: DATA CODE
- pandas: dtypes, missing values, groupby, merge, pivot
- One messy public dataset cleaned end to end

WEEKS 5-6: EVALUATION + ENGINEERING
- Metric choice per project, baselines, time/group splits
- requirements.txt, seeds, Pipeline, two pytest tests

WEEK 7 ONWARD: MOCK INTERVIEWS
- Two mocks a week, one theory and one project

Can I do this without notes?

[ ] Explain overfitting using my own project
[ ] Tell high bias from high variance by looking at scores
[ ] Clean a column with commas, spaces and missing values
[ ] Write a groupby with two aggregations
[ ] Pick and defend a metric for an imbalanced problem
[ ] Explain why my split matches real use
[ ] Show how to rerun my project from a fresh clone
[ ] Point to at least one test I wrote

Each unchecked box = one focused practice session.

WEAK
- Built a machine learning model for loan prediction
  with 92% accuracy using Random Forest.

STRONGER
- Predicted loan default on ~____ applications; chose recall
  as the main metric because missed defaults cost most.
- Beat a logistic regression baseline by ____ recall using
  a tuned random forest; validated on later applications.
- Packaged preprocessing in a scikit-learn Pipeline with
  pytest tests; project reruns from one command.

Fill the blanks with your own real numbers only.