Vitalii Ivanov

Senior Software Architect / Tech Lead

Helsinki

Here you'll find tutorials on technologies I recently found fascinating, or simply wanted to brush up on.

Tutorials

Showing all 27 tutorials

OAuth and JWT: how one sign-in becomes an API call

Tags:Other

The authorization endpoint and consent, the token endpoint and PKCE, JWTs and how an API checks them, OpenID Connect, refresh tokens and revocation, the backend-for-frontend pattern and DPoP, one part at a time, from one sign-in to an API call.

Async Python: how one thread checks 100 links

Tags:Python

Sockets and the selector, coroutines, the event loop and futures, tasks and TaskGroup, cancellation and timeouts, semaphores and connection pools, the default thread pool, the GIL, interpreter pools, free-threaded 3.15 and debug mode, built up one part at a time around one link checker.

AWS Bedrock and SageMaker: how an AI support chat is put together

Tags:AWSAI

Bedrock’s Converse API, inference profiles, the accounts Claude runs in, invocation logs, GPU instances, SageMaker AI endpoints, autoscaling, batch, PrivateLink, SCPs and guardrails, added one at a time until a support chat’s whole AI stack fits on one map.

Prometheus: how a monitoring stack is put together

Tags:Other

Client libraries, exporters, scraping, the Pushgateway, the TSDB, PromQL, dashboards, rules and Alertmanager, one part at a time, on a two-copy checkout service, with PCA exam cues.

AWS Route 53: how a company’s DNS is put together

Tags:AWS

Hosted zones, records, routing policies, health checks, the VPC Resolver, private zones, hybrid endpoints, DNS Firewall, DNSSEC, query logs and Profiles, added one at a time until a company’s whole DNS fits on one map.

Kubernetes: how a cluster is put together

Tags:Other

The API server, etcd, controllers, scheduler, kubelet, Services, CoreDNS, volumes and node heartbeats, one part at a time, on a three-node cluster running one small app, with KCNA exam cues.

AWS access: how one laptop reaches a private database

Tags:AWS

Access keys, SigV4, STS roles, the credential chain, IAM Identity Center, Session Manager and RDS IAM auth, one part at a time, on the path from one laptop to a private database.

AWS KMS: how encryption at rest is put together

Tags:AWS

KMS keys, data keys, key policies, grants, CloudTrail, cross-account and multi-Region keys, rotation, key stores and deletion, added one at a time until the whole setup fits on one map.

AWS VPC: how a private network is put together

Tags:AWS

Subnets, route tables, security groups, network ACLs, NAT, endpoints, PrivateLink and Transit Gateway, added one at a time until the whole network fits on one map.

AI benchmarks: what the numbers actually measure

Tags:AI

What one benchmark test looks like, who writes the tasks, who runs them, who grades the answers, and whether the model has already seen the questions.

The AI stack: how one claim crosses six layers

Tags:AI

LangChain, LangGraph, Langfuse, Airflow and the Claude Agent SDK: which job each one owns, shown on one insurance claim.

Diffusion models: how noise becomes a picture

Tags:AI

How a prompt and a grid of random numbers become an image, and why the network only ever predicts noise.

MCP: how one question becomes a tool call

Tags:AI

How one question travels from your editor to a tool and back: the protocol every AI client now speaks.

Claude Code: how one prompt becomes a code edit

Tags:AI

What a coding agent actually does between your prompt and the two lines it ends up changing.

Git worktrees: how three AI agents share one repo

Tags:AI

How several agents work in the same repository at once without overwriting each other.

Interpreted vs compiled: how one line reaches the CPU

Tags:Python

One line of code through Python, JavaScript, Java and Go, and why the split is a dial, not a dichotomy.

Reasoning models: how a model learns to think before it answers

Tags:AI

How DeepSeek-R1 and o1 learned to think before answering, followed through the AIME 2024 score from 9.3% to 93%: a prompt, RL with a 0-or-1 reward, a thinking budget, 64 tries, and what the hidden working costs.

Concurrency: how Java, Node, Go & Python hold 10,000 clients

Tags:Python

The same small server in four languages, hit by 10,000 clients at once: where each one keeps a waiting request, and why the answers took from 0.4 to 5.2 seconds.

Transformers: how five words become a sixth

Tags:AI

One five-word prompt followed through GPT-2 small: tokens, embeddings, attention, 12 layers and the final scores, and the 928 million operations it takes to pick the sixth word.

LoRA: how to teach a frozen model in 8 megabytes

Tags:AI

How to teach a frozen model a new skill by training a tiny patch instead of the weights.

Prompt injection: how a poisoned email hijacks an AI assistant

Tags:AI

One email nobody opens makes an AI assistant send 512 characters of customer data to a stranger. Follow it from inbox to leak, then see which defenses bring the count back to 0.

AI evals: how a one-line prompt change earns the right to ship

Tags:AI

One line is added to a support agent’s prompt. Code checks, an LLM judge, three reviewers, 200 agent runs, statistics and a canary decide whether it ships.

AI numbers every AI engineer should know

Tags:AI

The handful of numbers that explain what a model costs, how fast it answers, and why.

AI agent patterns: how one support ticket gets closed

Tags:AI

What an agent is made of, and every core pattern, shown on one support ticket with real code.

RAG: how one question finds its answer in 2 million words

Tags:AI

One question about the federal uniform allowance, answered from 2,034,544 words of US law: chunking, embeddings, cosine similarity, HNSW search, reranking and a cited answer, every number from one real run.

Neural networks: how three architectures read one handwritten 7

Tags:AI

Pixels, neurons, softmax, backpropagation, convolution and recurrence, followed through one MNIST seven and the 2,614,536 multiplications three networks spend reading it.

How a browser loads a page: from Enter to pixels

Tags:Other

One Enter key followed for 320 ms and 132 packets: the keypress, URL parsing, DNS, TCP, TLS 1.3 with a post-quantum key share, HTTP/2, Google’s front end and the render pipeline.