Domain 1: Computer Science & Software Engineering
Python from first principles to how CPython, the GIL and async really work, plus C++ memory, operating systems, networking, databases and SQL for ML.
Lessons
30 lessonsValues, Names and Types — What a Variable Really Is
Atlas stores each student's settings. One wrong assumption about what = does, and one student's settings leak into everyone else's.
Control Flow — Decisions, Loops and Loops That Know When to Stop
Atlas routes each question to the right model, walks through pages of results, and retries a flaky API without hammering it.
Functions — Arguments, Scope and the Call Stack
Atlas's code is growing. Functions turn copy-pasted logic into named, reusable, testable pieces, if you avoid the one default-argument trap that leaks data between users.
Lists and Tuples — Sequences, Copies and What They Cost
Atlas keeps a queue of questions waiting for the model. The wrong list operation turns that queue from instant into the slowest part of the system.
Dicts and Sets — Hashing and O(1) Lookups
Atlas checks 'have we seen this question before?' millions of times a day. With a list that check gets slower every day; with a set it stays instant.
Strings and Unicode — Text Is Harder Than It Looks
Every question Atlas answers is text: in Hindi, Tamil, emoji and English. Text that looks identical on screen can be different bytes, and that breaks search, dedup and billing.
How Python Really Runs Your Code
Atlas starts life as a Python program that stores lessons and answers questions.
The GIL — Threads, Processes and Which One to Use
Atlas preprocesses thousands of documents and calls the model API thousands of times. One job needs processes, the other needs threads or async, and mixing them up wastes the whole machine.
C++ and Memory — What Python Hides From You
The fastest parts of AI systems — inference engines, CUDA kernels, tokenizers — are written in C++ or similar languages. In C++ *you* decide where data lives and when it's freed.
Data Structures That Power AI Systems
Thousands of students use Atlas at once. Each request must find that student's history (lookup), wait its turn fairly (scheduling), and fetch the best-matching lessons (ranking).
Comprehensions, Iterators and Generators — Data Too Big for Memory
Atlas's training data is a 40 GB file of conversations. Reading it the obvious way crashes the machine; reading it one line at a time uses a few kilobytes.
Closures, Decorators and Context Managers — Reusable Behaviour
Every model call in Atlas needs timing, retries and caching. Written by hand, that's 20 lines around every call. Written once as decorators, it's one line on top of each function.
Classes, Dataclasses and Type Hints — Modelling Your Domain
Atlas passes requests between a dozen functions as raw dicts. One typo, 'max_token' instead of 'max_tokens', and the limit silently stops working. Classes and types turn that typo into an error you see before shipping.
Errors, Logging, Tests and Packaging — Code You Can Trust
Atlas is now real software with users. When it breaks at 2 a.m., good exceptions and logs tell you why in minutes; good tests stop it breaking in the first place.
NumPy and Pandas — Vectorized Python for Data and ML
Atlas scores 2 million question–answer pairs every night. A Python loop takes 40 minutes; the same maths as one NumPy expression takes under a second.
Profiling — Find the Slow Line Before You Fix Anything
Atlas's answer endpoint takes 1.8 seconds. Everyone has a theory: the model, the database, JSON. A profiler ends the debate in two minutes.
Async Python — Thousands of LLM Calls Without Thousands of Threads
Atlas grades 10,000 student answers with an LLM every evening. One call at a time takes 14 hours. Async with a concurrency limit takes 12 minutes, without tripping the provider's rate limit.
Interview Patterns and Capstone — An Async, Rate-Limited LLM Batch Client
Everything in the Python path comes together: Atlas gets a batch client that grades 10,000 answers safely, and you get a project and a set of patterns you can explain in any interview.
Operating Systems — Sharing One Machine Between Many Programs
Atlas runs a web server, background workers and a model on one machine. The operating system (OS) decides who runs, who gets memory, and how they communicate.
Networking — How a Question Travels to the Model and Back
Atlas has to be reachable from any browser. Every answer crosses the internet, so the network adds delay, can fail halfway, and can be overwhelmed.
Databases — Remembering Things Safely and Finding Them Fast
When a student finishes a lesson, closes the tab and returns next week, their progress must still be there — even if the server crashed in between, and even if two devices saved at the same moment.
SQL Basics — Tables, SELECT, WHERE, ORDER BY and the Truth About NULL
Atlas's analytics questions ('which students started a lesson but never finished it?') all start as SQL. Getting the basics exactly right, especially NULL, decides whether the answers are true.
SQL Joins — Inner, Left, Semi and Anti Joins, and Subqueries vs Joins
Every Atlas metric combines tables: students with enrolments, enrolments with lesson progress, progress with payments.
SQL Aggregation — GROUP BY, HAVING, Dates, Buckets and Pivots
Daily active students, completion rate per course, revenue per country per month: Atlas's dashboards are aggregations.
SQL Window Functions — Ranking, Running Totals, Top-N and Gaps and Islands
'Each student's three most recent lessons', 'running total of study minutes', 'longest daily streak': Atlas asks per-row questions that look at neighbouring rows.
SQL CTEs and Recursive Queries — Readable Pipelines, Hierarchies and Funnels
Atlas's lessons form a prerequisite graph, and its product questions are multi-step funnels (visited → started → finished → paid).
SQL Under the Hood — Query Plans, Indexes, Columnar Engines and Transactions
Atlas's progress dashboard got slower every week until it took 40 seconds. Nothing in the query had changed; the table had grown, and the database was reading all of it on every request.
SQL for ML — Point-in-Time Training Sets, A/B Test Analysis and Vector Search in Postgres
Atlas's dropout model looked brilliant offline (AUC 0.94) and useless in production. Its training features were computed with data from after the prediction date.
SQL Interview Patterns and Capstone — A Product Analytics Warehouse
Atlas's analytics grew as ad-hoc queries against the production database, each defining 'active student' slightly differently.
Bridge: From Storing Text to Understanding It
Atlas can store and serve everything — but understands nothing it stores.