Vitalii Ivanov

Coroutines, the event loop, tasks and TaskGroup, cancellation and timeouts, semaphores and queues, blocking calls and threads, the GIL, subinterpreters and free-threaded Python 3.15, followed through one link checker and the 494 switches its single thread makes.

Inference profiles and cross-Region routing, data residency, how Claude is hosted in AWS, GPU and accelerator instances, SageMaker AI endpoints, autoscaling, batch and PrivateLink, followed through one support chat and the $5,386 monthly bill its design adds up to.

Metrics, scraping, the time series database, PromQL, alerting rules and Alertmanager, followed through one bad deploy and the 402,329 samples Prometheus stores around it, with PCA exam cues.

Hosted zones, alias records, routing policies, health checks and failover, the VPC Resolver, hybrid and multi-account DNS, DNS Firewall and DNSSEC, followed through one checkout and the 16 DNS queries it causes.

The API server, etcd, controllers, scheduler, kubelet, kube-proxy and CoreDNS, followed through one app on a three-node cluster and the 220 writes it causes, with KCNA exam cues.

Access keys, session tokens, SigV4 signing, the credential chain, IAM Identity Center and Session Manager, followed through one laptop and the 21 requests it sends to AWS.

AWS owned, AWS managed and customer managed keys, envelope encryption, key policies, cross-account access, rotation and deletion, followed through one invoice and the 20 KMS calls made for it.

Subnets, route tables, security groups, NAT, endpoints and Transit Gateway, followed through one checkout request and the 57 checks it passes.

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.

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

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

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

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

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

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

What a model is really doing while it thinks, and why the thinking is what makes it right.

Four runtimes under the same flood of connections, and the two tricks all of them rely on.

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.

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

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.

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.

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

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

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.

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

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.