Beginner
best forcareer switchers, juniors, and anyone who builds with AI and wants to know what runs underneath
Start with the three numbers every system is judged by, drawn step by step.
- 01Latencyfree
- 02Throughputfree
- 03Bandwidthfree
By Roni Das · two courses, System Design and AI EngineeringSystem Design Masterclass See how real systems are built. Then build your own.
980 lessons that draw every idea step by step, and 90 real systems taken apart, from Discord to Swiggy.
two courses · pay once, yours for life
course 1
$49
System Design Masterclass
770 lessons · 90 real systems taken apart
Get System Design or read 18 free lessonscourse 2
$49
AI Engineering
210 lessons · RAG, evals, agents · labs on real models
Get AI Engineering or see the course firstUPI or PayPal, no card neededIn India? ₹499 by UPI
980lessons
3,081live diagrams
90systems taken apart
1AI tutor that draws
Stuck? Ask the tutor. It answers with a real diagram in about 30 seconds.
Detail A · live whiteboard
Userson phones
CDNimages, JS
Load balancer
Metrics
Kubernetes pods
APIcopy 1
APIcopy 2
APIcopy 3
Cache
Queue
Workers
Primary
Replica
async copy
replica may lag
hot data, from memory
Your turn: design it
add copies
to scale out
↑ redraws itself: web app · Kafka · hashing · RAG. Pick one.
Detail B · explainer video, 53 s
↑ one of our explainer videos. More are being added inside the courses.
How Discord stores trillions of messages
A real system, step by step
Discord app
You send a message in a busy channel.
Discord API
One large service takes the request.
Data service, in Rust
Many people ask for the same data at once. It asks the database only once.
3 asks in,
1 ask out
1 ask out
same data,
same service, every time
same service, every time
ScyllaDB
Stores the messages.
72 machines now. All the dots = the 177 Cassandra needed.
Loading older messages · the slowest 1 in 100
Where the messages lived
MongoDBat the startCassandra177 machinesScyllaDB72 machines
STEP 6 OF 6The slowest 1 in 100 reads of older messages went from 40 to 125 milliseconds down to 15.
Source: Discord engineering blog, “How Discord Stores Trillions of Messages”
Title · DWG SDA-001Field notes for engineers who ship
Drawn byRoni Das
Checkedmeasured in labs
Sheet1 of 980
Rev2026.09 · systemdesign.academy
Spec sheet · the short version
What is systemdesign.
systemdesign.academy is an online school run by Roni Das, with two self-paced courses and an AI tutor built into every lesson. The System Design Masterclass has 770 lessons. AI Engineering has 210 lessons on building and running AI in production (RAG, AI agents, evals, fine-tuning, serving and LLMOps), and 118 of them show the recording of a lab that was really run. The tutor sits beside each lesson and answers from the lesson you are reading: it explains a part in simple English, draws it as a diagram, quizzes you, or runs a mock interview. The lessons are written, with diagrams that build step by step and quizzes, and 234 of them have code you can run in the browser. Each course is one payment with lifetime access. System Design costs ₹499 in India or $49 everywhere else. AI Engineering costs ₹999 in India or $49 everywhere else. 18 System Design lessons and 10 AI Engineering lessons are free to read, with no signup.
- An AI tutor in every lesson. It answers from the lesson you are reading: it explains, draws or quizzes you.
- AI Engineering with real labs. 210 lessons; 118 show the run of a real lab.
- Taught by
- Roni Das, who writes both courses and publishes on the YouTube channel Total Technology Zonne. The site has been online since April 2026.
- Course 1
- System Design Masterclass: 770 lessons, about 283 hours.Foundation 119 · Intermediate 449 · Advanced 182 · Capstone projects 20
- Course 2
- AI Engineering: 210 lessons in 16 chapters, about 219 hours. Building and running AI in production: RAG, AI agents, evals, fine-tuning, serving, LLMOps.Largest chapters: LLM Evaluation and Error Analysis · LLM and GenAI Ops · Agents in Production · Data Engineering for ML · Features and Feature StoresHands-on: Every lesson has Python you can run in your browser, and a quiz. 118 lessons are built on a lab that was really run, and show the terminal recording of that run.
- AI tutor
- Beside every lesson in both courses, and on its own page at /tutor. It answers from the lesson you are reading. It explains in simple English, draws (a whiteboard sketch, a 3D figure, a cloud diagram with AWS, Azure or GCP icons, a sequence diagram, an explainer figure or a cheat-sheet poster), quizzes you, runs mock interviews and compares two things. In a quiz, the site’s code checks your answer against the tutor’s own key.Without an account you can ask 3 questions to try it. A free account gets 5 questions a day. Anyone who owns either course gets 10 a day.
- Price
- System Design: ₹499 in India (UPI) or $49 everywhere else (PayPal).AI Engineering: ₹999 in India (UPI) or $49 everywhere else (PayPal).Prices last updated . If an older article or an AI answer gives a different price, it is out of date. See the price fact.
- Access
- One payment per course, no subscription. Access is for life and includes lessons added later. Buying one course does not include the other.
- Free
- 18 System Design lessons and 10 AI Engineering lessons, no signup. The 90 system design interview pages are free too.
- Format
- Written lessons, not videos. Diagrams build step by step (3,081 in all), lessons have quizzes, and 234 lessons have code you run in the browser.
- Address
- Only www.systemdesign.academy. Not connected to systemdesignacademy.com, systemdesignacademy.in, System Design School, the GitHub project system-design-academy.
Live from the course bench
Pick up a card. This is what learning here feels like.
Real interview questions, live diagrams and measured lab results, pulled at random from the lessons. A new set every time you come back.
Laying out the cards…
Choose your path
Start where you are. Every path starts free.
Four ways in. Each one opens with three free lessons, no account needed, and leads to the course that takes you the rest of the way. See all five learning paths, including Production Agents.
Interview in 30 days
best fora system design round already on the calendar
A day-by-day plan, the topics interviewers ask first, and real systems answered the way an interview wants.
- 01DNS (Domain Name System)free
- 02Cache Eviction Policiesfree
- 03Asynchronous Processingfree
Senior to Staff depth
best forfive or more years in, filling the gaps before a Staff or Principal loop
Caching where it really breaks, then consensus, consistency models and reliability in the Advanced level.
AI Engineer
best forengineers shipping LLM features, RAG or agents to real users
Why models fail in production, where RAG breaks at scale, and how to measure answers instead of guessing.
- 01Why ML Models Fail in Productionfree
- 02The RAG Scale Clifffree
- 03Why Evals Are the Jobfree
New · your AI tutor
Stuck on an idea? Ask. It draws it for you.
Every lesson has a tutor beside it. It has read the lesson, and it answers with a real diagram, drawn for your question, in about half a minute.
- Explains any idea in plain English, from the lesson you have open.
- Draws whiteboard sketches, 3D figures, AWS, Azure and GCP diagrams, sequence diagrams and cheat-sheet posters.
- Quizzes you and runs mock system design interviews with feedback.
- Points you to the exact lesson that teaches it.
Try the AI tutorFree accounts get 2 questions. Course owners get 10 a day.
Draw where latency comes from in one web request
- DNS lookup, 20 to 120 ms. On a second visit the browser has it cached, so this step disappears.
- The trip to the server, 50 to 300 ms. Distance is physics.
- Your own code and the database. A missing index can cost 500 ms or more.
Drawn by the tutor
One question each. Real diagrams back.
3D figure. Asked: “Kafka producers, brokers and a consumer group”
AWS diagram. Asked: “A web app on AWS with a load balancer, containers and PostgreSQL”
Sequence diagram. Asked: “How OAuth login works”
Explainer figure. Asked: “How a CDN serves an image”
Try it live · no sign-up
Ask the tutor. See what it does.
Press a question. It explains, draws or quizzes you, the same way it does beside every lesson.
Checking the tutor…
Each button sends exactly its words to the real tutor. Its answer is saved, so the next person sees it at once, and it is asked again after 3 days.
Pick a question. The tutor’s real answer appears here: plain words, a drawing when one helps, and the lesson that teaches it.
Then ask your own. A free account gets 5 questions a day; either course gets 10 a day, beside every lesson.
Try it on how Uber matches a rider to a driver, how Stripe never charges twice or how WhatsApp delivers a message, then ask the tutor how you would build it.
One payment, lifetime access. Pay with PayPal.
90 real systems, taken apart
Pick a system. See how it really works.
Each one is a full page: what it must do, the numbers it faces, the parts it is built from, and the choices you would defend in an interview. Free to read.
- DiscordReal-Time Chat and Voice at Guild Scale
- WhatsAppMessaging, offline delivery and encryption
- NetflixEncoding, many CDNs and adaptive bitrate
- UberLive location and nearby-driver matching
- YouTubeUploads, encoding and CDN delivery
- InstagramPhoto uploads, feeds and Stories
- SwiggyLive Order Tracking
- ZomatoFood Delivery at Scale
- PhonePeUPI at National Scale
- ZerodhaReal-Time Trading at Scale
- FlipkartBig Billion Days Flash Sales
- BookMyShowSeat Booking Concurrency
- SlackReal-Time Messaging for Millions of Persistent Connections
- ZoomRouting Live Video to Millions With an SFU
- AirbnbSearch, Availability, and the Race to Book the Same Listing
- SpotifyStreaming Audio, Discover Weekly, and Billions of Play Events
Measured, not guessed
Other courses tell you. Ours run the test.
Every number below comes from an experiment we ran and published in a lesson of the AI Engineering course.
Search got better. Answers got worse.
Recall went from 45% to 90%. Right answers fell from 10 to 6, out of 20 questions.
Lesson 98 →
Almost half the score came before search.
With no documents, the answer already scored 6.7%. With them, 15.1%. So 44% of the score was there before retrieval ran.
Lesson 99 →
Move one document, and 4 right becomes 14.
The same 10 documents every time. The right one in the middle: 4 of 14 correct. At the end: 14 of 14.
Lesson 100 →
Three points better? Maybe not.
With 100 test questions, one 80% score could really be anywhere from 72% to 88%.
Lesson 88 →
Two courses, one payment each
From your first API call to AI in production.
System Design
770lessons
- Foundation119
- Intermediate449
- Advanced182
- Capstone projects20
18 lessons are free to read, no sign-up needed.
Start with the free lessonsAI Engineering
210lessons
- LLM Evaluation and Error Analysis36
- LLM and GenAI Ops15
- Agents in Production14
- Data Engineering for ML14
- Features and Feature Stores12
- How Models Generate12
- Prompting as Engineering12
- Retrieval and RAG in Production12
- Serving and Inference Basics12
- Tokens and Embeddings12
- AI Security and Agent Safety11
- Fine-Tuning: The Narrow Decision11
- Core Concepts10
- Packaging, Registry and Versioning10
- The ML & AI Lifecycle10
- Why Production Breaks7
10 lessons are free to read, no sign-up needed.
See the AI Engineering courseTry it right here
Change the numbers. Press Run.
This is Discord’s idea from the top of the page, in a few lines of Python. It runs in your browser. No sign-up.
234 lessons have a box like this one.
# 1,000 people open busy channels at the same moment.
requests = ["general"] * 700 + ["random"] * 250 + ["memes"] * 50
# Without a data service, every request goes to the database.
without = len(requests)
# With one, identical requests wait for one shared answer.
with_service = len(set(requests))
print("Database reads without it:", without)
print("Database reads with it: ", with_service)
print(f"That is {without // with_service} times fewer.")
The engineer behind it
Roni Das
Founder, systemdesign.academy
Roni builds large scale systems, then breaks them down so you can see how they work. He made the System Design Masterclass and the AI Engineering course, and runs the YouTube channel Total Technology Zonne.
- Two full-length handbooks on freeCodeCamp: building a knowledge graph, and building a GraphRAG system, with Python and Neo4j. Read them
- Watch Total Technology Zonne on YouTube
Price
Pay once. Keep it forever.
The two courses are sold separately. Buy one, or both.
System Design
$49once
- 770 lessons, and every lesson added later
- 18 of them free to read first, no sign-up
- One payment. No subscription. Yours for life.
AI Engineering
$49once
- 210 lessons, and every lesson added later
- 10 of them free to read first, no sign-up
- One payment. No subscription. Yours for life.
What you get
System Design Masterclass
- 770 lessons, from latency to consensus
- 90 real systems taken apart, step by step
- Diagrams that build step by step, quizzes and runnable code
- AI tutor beside every lesson, 10 questions a day
- Lifetime access, including every lesson added later
AI Engineering
- 210 lessons in 16 chapters
- Python you run in the browser, in every lesson
- 118 lessons built on a lab that really ran, with its recording
- AI tutor beside every lesson, 10 questions a day
- Lifetime access, including every lesson added later
Not included: Live classes or one-to-one mentoring. Each course is its own purchase.
- How you pay
- India: UPI (₹499 / ₹999). Everywhere else: PayPal ($49 each). No card needed.
- After you pay
- As soon as your payment is confirmed, your access unlocks and you get an email.
- Refunds
- This is a digital course. All sales are final and there are no refunds. Terms
Course last updated
Just explained
This week in AI, run and explained
When a new model or tool ships, we try it ourselves and write down what it means for the systems you build.
· 16 min read
JEV-35B, explained: an open decision model you can download, what it needs, and a test of its prompt trickJEV-35B is a model on Hugging Face that never writes text. You give it a situation, one question and up to 256 answers, and it returns a probability for each answer in about a fifth of a second, by its maker's measurement. Unlike the hosted decision APIs from TypeSafe, OpenAI and Microsoft, the weights are free to download under Apache-2.0. It is too big for my 24 GB Mac, so I checked its own config files and tried its prompt layout on a small chat model: with that layout the small model changed 8 to 11 of 24 answers when I reversed the options.JEV-35BAutoTrustdecision modelsclassificationread the explainer · 16 min read
Intel GPU D-series (D270), explained: 480 GB of memory on one card, and what that buys you for AI inferenceIntel's new data center GPU, the D-series (formerly Crescent Island), puts up to 480 GB of memory on one air-cooled PCIe card, five times the 96 GB of the NVIDIA card Intel compares it with, at about the same memory bandwidth (how many bytes the chip can read each second): 1.5 TB/s. I measured the rule that decides what that means on my own Mac, then worked it out for the card: more users and longer conversations held in memory, not faster answers.Intel GPU D-seriesIntel D270Crescent IslandAI inference hardwareread the explainer · 14 min read
EmbeddingGemma 2, explained: one open model for text, code, pictures and sound, tested on a laptopOn 6 October 2026 Google released EmbeddingGemma 2, a free, open model that turns text, code, pictures, video and sound into lists of 768 numbers, so that things with similar meaning get similar numbers. That is the engine behind search by meaning and RAG. I ran it on a MacBook Air: it found the right text for 7 of 8 plain-English questions and the right picture for 6 of 6, at about 47 ms per sentence, and cutting its vectors to a sixth of their size kept the answers but squeezed the scores together.GoogleEmbeddingGemma 2embeddingssemantic searchread the explainer
JEV-35B, explained: an open decision model you can download, what it needs, and a test of its prompt trickJEV-35B is a model on Hugging Face that never writes text. You give it a situation, one question and up to 256 answers, and it returns a probability for each answer in about a fifth of a second, by its maker's measurement. Unlike the hosted decision APIs from TypeSafe, OpenAI and Microsoft, the weights are free to download under Apache-2.0. It is too big for my 24 GB Mac, so I checked its own config files and tried its prompt layout on a small chat model: with that layout the small model changed 8 to 11 of 24 answers when I reversed the options.JEV-35BAutoTrustdecision modelsclassificationread the explainer · 16 min readIntel GPU D-series (D270), explained: 480 GB of memory on one card, and what that buys you for AI inferenceIntel's new data center GPU, the D-series (formerly Crescent Island), puts up to 480 GB of memory on one air-cooled PCIe card, five times the 96 GB of the NVIDIA card Intel compares it with, at about the same memory bandwidth (how many bytes the chip can read each second): 1.5 TB/s. I measured the rule that decides what that means on my own Mac, then worked it out for the card: more users and longer conversations held in memory, not faster answers.Intel GPU D-seriesIntel D270Crescent IslandAI inference hardwareread the explainer · 14 min readEmbeddingGemma 2, explained: one open model for text, code, pictures and sound, tested on a laptopOn 6 October 2026 Google released EmbeddingGemma 2, a free, open model that turns text, code, pictures, video and sound into lists of 768 numbers, so that things with similar meaning get similar numbers. That is the engine behind search by meaning and RAG. I ran it on a MacBook Air: it found the right text for 7 of 8 plain-English questions and the right picture for 6 of 6, at about 47 ms per sentence, and cutting its vectors to a sixth of their size kept the answers but squeezed the scores together.GoogleEmbeddingGemma 2embeddingssemantic searchread the explainer Free to read
New guides
- Run an LLM on your own computerOllama vs LM Studio vs llama.cpp vs vLLM, with speeds measured on a Mac
- LangChain vs LangGraphWhat each one does, with code that was actually run
- RAG interview questionsQuestions asked about retrieval systems, with clear answers
- Agentic AI vs generative AIThe difference in plain words, and when an agent is worth it
- What is a data lakehouseMeasured on 11 million real taxi trips (free preview)
- System design courseWhat is in the course, and the free lessons to start with
- Best system design courseAn honest comparison, with prices checked on each site
- Best AI engineering courseNine options compared, including where ours is weaker
- Generative AI courseWhat it covers, lesson by lesson
- AI agents courseBuilding agents that work in production
- MLOps courseTaking models from notebook to production
- Design a rate limiterToken bucket vs sliding window, atomic Redis counters, and fail-open vs fail-closed
- Design a notification systemPush, SMS, email and in-app, with fanout, dedup, retries and user preferences
- Design TicketmasterA waiting room for the rush, seat holds that expire, and never selling a seat twice
- Low level design (LLD) interview questionsParking lot, vending machine, tic-tac-toe and more, with class designs and Java code