Reference

Glossary

The canonical terms for this course. Lessons use these words with these meanings.

Roles

AI engineer
Builds applications on top of existing foundation models (LLMs) by adapting them with prompting, RAG, agents and fine-tuning. Evaluation is central to the work. Huyen
ML engineer
Trains and ships models: builds the model first, then the application around it. In AI engineering the order is usually reversed. Huyen
Evaluation
Judging whether a model's (or an AI assistant's) output is correct and fit for purpose. The skill that becomes more valuable as AI writes more of the code.

Habits

Predict → Run → Compare
Before running AI-written code, write down what it will return. Then run it and compare. When your prediction was wrong, ask why. Lesson 01

pandas

DataFrame
A table of rows × columns. pd.DataFrame.
Series
A single column. It works like a hashmap whose keys are the index labels.
NaN
"Not a Number". pandas uses it to mark a missing value. Aggregations skip it by default, and any comparison with it returns False.
Boolean mask
A Series of True/False values, such as df["sales"] > 100. Using it to index a table (df[mask]) keeps only the rows where it's True.
groupby
Split → apply → combine: split the rows by a key, calculate something (an aggregation) for each group, then combine the results. Groups come back sorted by key.
Aggregation
Reducing many values to one, e.g. sum, mean, count.
Share (proportion)
The fraction of rows that meet a condition. In pandas it's the mean of a True/False column, because True counts as 1 and False as 0.
datetime
pandas' type for real dates (datetime64). A date stored as text looks the same when printed but can't be compared, grouped by month or used with .dt.
Period
A whole span of time, like the month 2026-03, rather than one moment. .dt.to_period("M") turns dates into months so you can group by them.
Rolling average
An average over a sliding window, e.g. this month and the two before. It smooths a noisy line so one unusual month doesn't dominate.
Sample size
How many values an average is based on. A 5.0 average from one review tells you far less than a 4.2 from two hundred, so always show the count next to the mean.