How a web request works
Follow one request from the browser through DNS, the internet, a CDN, a load balancer and an app server to the database and back.
2:21 · 10 chapters
Animated explanations
71 concept videos, each a live Remotion composition built from the same component library as the lessons: same colours, same icons, same mental model. Use chapters to jump, or pause at each step. The short story videos inside each lesson section are not listed here; watch them in their lesson.
Client vs server
Frontend vs backend: what runs in the browser, what runs on your servers, and why the server must never trust the client.
1:42 · 7 chapters
Anatomy of an API call
A hotel Booking API end to end: GET and POST, headers and JSON bodies, status codes, error responses and safe retries with idempotency keys.
1:41 · 7 chapters
The database bottleneck
Watch a database collapse as traffic grows from 1,000 to 100,000 requests per second, then fix it step by step with indexes, caching, read replicas and connection pooling.
2:17 · 10 chapters
Performance vs scalability
Making one request faster versus keeping latency flat as traffic grows, with a latency-vs-load curve for a scalable and an unscalable system.
1:28 · 6 chapters
From monolith to microservices
A growing team outgrows one codebase: modular monolith first, then independently deployed services with their own databases, and the real cost of going distributed.
1:46 · 7 chapters
Docker and containers
Why code that works on one machine fails on another, and how a Docker image packages runtime, libraries, dependencies, code and config so the same container runs everywhere.
1:52 · 8 chapters
Kubernetes: self-healing and autoscaling
Desired state vs actual state: a crashed pod is replaced automatically, then the Horizontal Pod Autoscaler scales from 3 to 10 pods under load and back down.
1:52 · 8 chapters
Event-driven architecture with Kafka
A BookingCreated event flows through a partitioned Kafka topic to five independent consumer groups; one crashes and replays from its offset while bookings keep working.
1:54 · 8 chapters
What happens in an LLM request
From prompt assembly to tokenization, next-token probabilities, the autoregressive loop, streaming, token cost and statelessness.
2:00 · 8 chapters
The agent loop
Watch an AI agent book a hotel through repeated think → act → observe cycles, with a human confirmation before the booking.
2:31 · 11 chapters
How tool calling works
Tools are JSON schemas, the model emits tool_use requests, and your code validates, executes and confirms side effects before returning tool_result.
1:46 · 8 chapters
RAG: retrieval-augmented generation
Index documents into a vector database, then embed a question, retrieve the nearest chunks and let the LLM answer from them with citations.
1:58 · 9 chapters
Embeddings and similarity
Words become vectors in a space of meaning: clusters, a query landing near related items, cosine similarity as an angle, and the king − man + woman ≈ queen intuition.
1:30 · 7 chapters
MCP: the Model Context Protocol
From N×M custom integrations to one standard plug: MCP hosts, clients and servers, and the JSON-RPC messages initialize, tools/list and tools/call.
1:53 · 8 chapters
Evals: testing AI systems
Build a test set, run prompt variants, grade with exact checks and an LLM judge, compare on a scoreboard and catch regressions on every change.
1:46 · 8 chapters
Prompt injection and defense in depth
An agent reads a web page with hidden instructions and leaks data, then the same attack fails against isolation, least privilege, a tool allowlist, output validation and human confirmation.
2:30 · 11 chapters
Model routing and AI gateways
An AI gateway routes simple requests to a small model, complex ones to a large model, images and audio to specialist models, compares costs and fails over on errors.
2:00 · 9 chapters
How the internet moves packets
Follow a request as packets with IP addresses, hop by hop through routers, and watch DNS resolve a hotel's domain through root, TLD and authoritative servers, with caching and TTL.
1:14 · 9 chapters
Inside one machine: processes and threads
Processes with isolated memory, threads sharing memory (and racing), CPU cores and the scheduler, context switches, and the latency ladder from L1 cache to the network.
1:15 · 7 chapters
The JavaScript event loop
Call stack, Web APIs / libuv, microtask and task queues: why one JavaScript thread handles many requests, how one blocking CPU task freezes everything, and how a worker fixes it.
1:13 · 8 chapters
How browsers render pages
HTML → DOM, CSS → CSSOM, render tree, layout, paint and composite; then CSR vs SSR vs SSG timelines showing what the guest sees when, and what hydration does.
1:14 · 8 chapters
Sessions, JWTs and OAuth
Server-side sessions with cookies vs signed JWTs (and why JWTs are hard to revoke), then "Sign in with Google" for a hotel app: OAuth 2.0 authorization code flow with PKCE and the OpenID Connect id_token.
1:12 · 7 chapters
Background jobs and webhooks
Move slow work out of the request: enqueue jobs, workers with retries and exponential backoff, a dead-letter queue, and inbound payment webhooks verified by signature and processed idempotently.
1:12 · 8 chapters
Transactions, locks and deadlocks
Booking + payment in one ACID transaction, the lost-update race when two guests book the last room, SELECT … FOR UPDATE row locks, and a deadlock the database detects and resolves.
1:11 · 7 chapters
Caching strategies
Cache-aside, read-through, write-through and write-back with a hotel price cache; TTLs and invalidation; and a cache stampede on a hot key, prevented with locking, early refresh and jitter.
1:13 · 8 chapters
TCP, UDP and TLS
The TCP three-way handshake, sequence numbers and ACKs, retransmission after loss, congestion window growth, UDP's fire-and-forget video call, and the TLS handshake on top.
1:09 · 7 chapters
CAP: partitions and trade-offs
Two replicas of a hotel's inventory in Paris and Lisbon lose their link. Choose consistency or availability, reconcile afterwards, and see the PACELC latency trade-off.
1:13 · 7 chapters
Retries, idempotency and circuit breakers
A payment call times out: naive retries double-charge the guest and cause a retry storm. Fix it with idempotency keys, exponential backoff with jitter, and a circuit breaker.
1:14 · 7 chapters
The saga pattern
A hotel booking spans Inventory, Payment, Loyalty and Email. When step 3 fails, compensating actions undo earlier steps in reverse. Then orchestration vs choreography.
0:59 · 6 chapters
Replication and sharding
Leader-follower replication, replication lag and read-your-writes, then sharding by hotel_id, a hot shard, and consistent hashing for resharding.
1:12 · 7 chapters
From 100 users to 100 million
One booking app grows step by step: separate database, load balancer, cache, read replicas, CDN, queues, shards and multiple regions, each forced by a visible bottleneck.
1:12 · 10 chapters
CI/CD and safe rollouts
A commit flows through CI (build, test, lint, security scan) into an image, staging and production. Then a bad release under rolling, blue-green and canary rollouts, and a feature-flag kill switch.
1:13 · 7 chapters
Cloud networking: regions, VPCs, subnets
Region and availability zones, a VPC with public and private subnets, internet and NAT gateways, security groups, and where the load balancer, app servers and managed database live.
1:13 · 7 chapters
Logs, metrics and traces
Debug one slow checkout: a metrics dashboard shows the p99 spike, logs reveal connection-pool timeouts, and a distributed trace pinpoints the slow database span. Then alert on SLO burn rate and ship it all with OpenTelemetry.
1:13 · 8 chapters
Failover and disaster recovery
A primary database dies: health checks detect it, the standby is promoted and the app reconnects. Then backups and point-in-time restore, RPO vs RTO on a timeline, and what multi-AZ and multi-region disaster recovery cost.
1:15 · 8 chapters
Injection, XSS and CSRF
Three OWASP classics shown safely: SQL injection vs parameterized queries, stored XSS in a hotel review vs output encoding and CSP, and CSRF vs SameSite cookies and CSRF tokens.
1:14 · 8 chapters
The testing pyramid
Unit, integration, contract and end-to-end tests compared on speed, cost and confidence, following a hotel price bug to the level that catches it, a flaky browser test, and CI running everything on every commit.
1:12 · 8 chapters
The system design method
Design a URL shortener step by step: requirements, back-of-the-envelope estimates (QPS, storage, 100:1 reads), API, data model, high-level design, bottlenecks (hot-link cache, ID generation) and trade-offs.
1:08 · 8 chapters
What architecture costs
Two architectures for the same traffic: a €1k/month monolith with managed Postgres and a CDN vs a €30k/month Kubernetes, microservices, Kafka, multi-region setup. Stacked cost bars, and when the expensive one is justified or waste.
1:04 · 7 chapters
Transformers and attention
From learned weights to tokens, embeddings, attention and stacked layers: how a transformer predicts the next token.
1:14 · 8 chapters
Context engineering
Treat the context window as a budget: what goes in, what stays out, what gets retrieved or summarized, and what it costs.
1:10 · 7 chapters
Agent memory
Short-term context vs a long-term memory store: extracting, summarizing and retrieving memories, plus stale, wrong and private memories and user control.
1:11 · 7 chapters
Voice AI pipeline
Microphone to speaker through speech recognition, an LLM, tools and text-to-speech; latency budgets, streaming, turn detection, barge-in, and realtime speech-to-speech models.
1:10 · 6 chapters
Multimodal AI
A hotel agent uses a vision model on room photos to open tickets, and document understanding on invoices with schema validation and human review.
1:00 · 6 chapters
Production agent runtime
User → AI gateway → agent runtime (model, memory, RAG, tools, policies, observability) → hotel systems, with one request traced and a policy blocking a risky payment.
1:09 · 7 chapters
AI cost optimization
LLM cost = requests × tokens × price. Apply prompt caching, model routing, context trimming, batching and a semantic cache, and watch the bill fall.
1:10 · 8 chapters
The product engineering loop
From a real hotel-guest problem to requirements, MVP, shipping, metrics, A/B experiments and iteration, with deliberate technical debt and no imagined scale.
1:13 · 8 chapters
Types, functions and the boundary
Variables and types, a text-plus-number checkout bug, static typing catching it early, pure pricing functions, and validating outside data at the boundary.
1:09 · 7 chapters
Data structures: pick by the job
Scanning a list vs jumping into a hash map, sorted trees for price ranges, queues and stacks, and graphs of places, all on hotel data.
1:07 · 7 chapters
Big O intuition
Steps for 1,000 vs 1,000,000 items for O(1), O(log n), O(n) and O(n²), the growth curves side by side, and replacing a search-inside-a-loop with a Set.
1:12 · 7 chapters
Where should this state live?
A hotel search page sorted into local, shared, URL, form and server state, including a stale 'last room' copy refreshed by a server-state cache.
1:02 · 7 chapters
Accessibility as engineering
The accessibility tree (role, name, state), real buttons vs clickable divs, keyboard focus, labels and non-colour errors, and ARIA used sparingly.
1:07 · 7 chapters
SQL and relations
From one duplicated sheet to guests, rooms and bookings tables: primary and foreign keys, a JOIN for a confirmation, and a CHECK constraint.
1:08 · 7 chapters
The NoSQL landscape
Key-value, document, wide-column, graph, search and vector stores, each fast at one access pattern, fed as copies from a relational source of truth.
1:12 · 7 chapters
Hexagonal and clean architecture
Classic layers vs a business core with ports and adapters: dependencies point inward, so payment providers swap and tests use fakes.
1:10 · 7 chapters
Serverless functions
An always-on thumbnail server sits idle most of the day. A serverless function starts on each photo upload, has a cold start, scales out for a burst and back to zero, and fits spiky rather than steady traffic.
1:15 · 7 chapters
Architecture Decision Records
Why decision reasons get lost, how a one-page ADR (context, options, decision, consequences) is written and reviewed in the repo, read later by a new engineer, and superseded instead of edited.
1:05 · 7 chapters
Designing a chat system
Why polling fails, how WebSockets plus a connection registry route messages, why the server stores and numbers a message before the first tick, how delivery acks give two ticks, how offline users catch up, and how message ids prevent duplicates.
1:09 · 7 chapters
Git, branches and pull requests
Commits as snapshots, a branch as a movable pointer, a pull request with tests and review, a same-day merge, and why a three-week branch explodes into conflicts.
1:05 · 7 chapters
Infrastructure as code
Click-ops drift versus a declarative file in Git: plan shows creates/changes/destroys, review and apply update the cloud and the state, the next plan catches manual drift, and the same files rebuild identical environments.
1:08 · 7 chapters
Docker volumes and networks
Uploads written inside a container vanish on redeploy; a named volume keeps database data across container replacement; localhost inside a container is the container itself; a user-defined network lets containers call each other by name; only the API port is published.
1:00 · 6 chapters
Docker Compose
From a stale 23-step setup guide to one compose.yaml: docker compose up creates a network, waits for Postgres to be healthy, runs migrations, starts the API and worker, which reach each other by service name. Compose is single-host, not a production orchestrator.
1:01 · 6 chapters
Kubernetes config and secrets
Why settings must not be baked into images; one image plus per-environment ConfigMaps; Secrets are base64, not encryption; encryption at rest and RBAC; env vars vs mounted files on rotation; external secret managers.
1:14 · 7 chapters
SLI, SLO, SLA and error budgets
Booking requests become an SLI, a goal (SLO) and a looser customer promise (SLA) on one ruler; the gap to 100% becomes an error budget that incidents spend, deciding whether to ship or to fix reliability.
1:10 · 7 chapters
The nines: downtime per year
A year of days shows what 99% really means; 99% to 99.999% become days, hours and minutes of downtime; a checkout chain multiplies availabilities, a spare server raises them, and each nine costs more.
1:11 · 7 chapters
Authentication vs authorization
A hotel receptionist logs in (authentication), then each request is checked (authorization): 401 vs 403, a refund over the limit, an IDOR on another hotel's booking and its server-side fix, and least privilege.
1:11 · 7 chapters
Secrets, hashing and encryption
Encoding vs hashing vs encryption decided by one question, a slow salted password hash checking guesses, a passport number encrypted with a KMS-held key, app secrets fetched from a secrets manager, and TLS plus encryption at rest.
1:14 · 7 chapters
Choosing a vector database
All vector databases find nearest neighbours; pgvector keeps vectors in the same Postgres row, a dedicated store needs a sync pipeline that can silently drop deletes; decide with numbers and move when memory or load forces you.
1:14 · 7 chapters
Hybrid search and reranking
A kiosk error code question: vector search buries the exact code, BM25 finds it, reciprocal rank fusion merges both lists into a shortlist, a cross-encoder reranker moves the right chunk to the top, and only the best few go to the model.
1:07 · 7 chapters
Build vs buy
A hotel platform's needs sorted into core (build) and context (buy), a three-year cost comparison for login, a vendor hidden behind an adapter and swapped when prices rise, and a written decision with a revisit trigger.
1:07 · 6 chapters