AI in Python — NumPy & Pandas, an interactive guide

01

Why NumPy?

Real example · payroll

Same task, two ways. Watch the clock.

You run payroll for a company and need to give everyone a 10% raise. A plain Python list does it one value at a time, in a loop. A NumPy array does it to the whole thing at once — that's called vectorization.

🏭

Why so much faster? A Python list gift-wraps every value in its own box and checks each one before touching it. A NumPy array packs identical values tightly together and processes them in one pass — an assembly line instead of inspecting items by hand.

Live: give 24 employees a 10% raise

0.000 sim seconds

Pick a method above

Animation is illustrative — the real gap on a million rows is roughly 50–100×.

02

The array: 1D, 2D, 3D

An array is just numbers arranged in a shape.

📦

1D is a single row of values (a line). 2D is rows and columns (a grid, like a spreadsheet). 3D is a stack of those grids (several sheets in one workbook).

Toggle the dimension

03

Inspecting an array

Four questions you ask about any array.

Build a grid by changing rows and columns, and watch the four attributes update. shape is how big, ndim is how many dimensions — they're easy to mix up.

Resize the grid

rows cols

04

Creating arrays & random

You rarely type arrays by hand. Generate them.

arange excludes the stop value and you set the step; linspace includes both ends and you set the count. rand gives floats 0–1; randint gives whole numbers — reroll them and watch.

Pick a generator

05

Indexing & slicing

Signature

Pick seats by position. The most-used skill in the language.

🎟️

Think numbered cinema seats. From the left, the first seat is 0. From the right, the last seat is -1 (there's no -0, which is why negatives start at -1). A slice a[start:stop:step] never includes stop.

Slice playground — type values or tap a preset

a[ : : ]

06

Reshaping

Signature

Same elements, new arrangement.

🥚

You bought 12 eggs. You can repack the same 12 into a 2×6 carton, 3×4, or one row of 12 — but never into a box that holds 10 or 15. Reshape works only when the element count stays the same.

Repack 12 elements

07

Operations & the axis

Signature

Aggregate down columns or across rows.

🧭

axis=0 points down the columns (collapses the rows) → a sum per column. axis=1 points across the rows (collapses the columns) → a sum per row. Axis 0 is vertical, axis 1 is horizontal.

Sum the grid

08

NumPy → AI

Real example · images

What a model actually "sees."

A tensor in PyTorch or TensorFlow is a NumPy-style array. An image is a 3D array: height × width × three color channels (R, G, B). Training is the operations you just saw — multiply, add, aggregate — run across these arrays millions of times. Hover a pixel to see its numbers.

An 8×8 image is just numbers

Hover any pixel…

Shape would be (8, 8, 3) — 8 rows, 8 columns, 3 color values each. That's 192 numbers for this tiny image.

09

The DataFrame

A spreadsheet you can program.

🚗

NumPy is the engine — raw numerical horsepower. Pandas is the car built around it: named columns, a row index, and mixed types (numbers and text and dates) so you can clean and analyze real-world data.

A DataFrame of employees

df = pd.DataFrame({"name", "age", "salary", "team"})

10

First look at any data

Three questions, three one-liners.

Before analyzing anything, ask: What's the shape? Is it clean? What do the numbers look like?

Inspect the employee data

11

.loc vs .iloc

Often confused

Labels are inclusive. Positions are not.

.loc selects by label and the range is inclusive. .iloc selects by integer position and the range is exclusive (like normal Python slicing). The i is for integer.

Select rows 0 to 2

stop value

df.loc[0:2] # inclusive

df.iloc[0:2] # exclusive

12

Filtering rows

Real example · movies

Keep only the rows that match a condition.

Drag the rating and pick a genre. The DataFrame filters live — exactly how you'd find, say, every highly-rated sci-fi film in a catalogue.

Filter the movie catalogue

rating ≥ 0 genre

13

GroupBy

Split into piles, summarize each pile.

🧺

GroupBy is sorting laundry into piles by color, then doing one thing to each pile. Here: group the movies by genre, then average the rating in each group.

Average rating by genre

14

Merging & joins

Signature · interview favorite

Combine two tables on a shared key.

Two tables share an id. The join type decides which rows survive — and where you get missing values (NaN).

Pick a join type

df1

df2

★

Cheat sheet

NumPy

np.array([1,2,3]) np.arange(0,10,2) · linspace(0,1,5) np.zeros/ones/full/eye np.random.rand() · randint() arr.shape .size .dtype .ndim arr.reshape(2,3) arr[1:4] · arr[::2] · arr[-1] np.sum(a, axis=0|1)

Pandas

pd.DataFrame({...}) df.head() .info() .describe() df["col"] · df[["a","b"]] df.loc[] · df.iloc[] df[df["age"] > 18] df.groupby("g")["x"].mean() pd.merge(a,b,on="id",how="left") df.to_csv() · pd.read_csv()

The three traps

arange excludes stop; linspace includes it axis=0 ↓ columns · axis=1 → rows .loc inclusive · .iloc exclusive

Master these three and most confusion disappears.