Journal

Blog

Notes from a running system: architecture, token economics, autonomy. They ship as things prove themselves in practice.

2026-07-11

Token economy: an LLM infrastructure for under three dollars a week

About 4,800 calls a week, a $2.40 bill. How an infrastructure works where a local model takes 92% of the load by volume and only the hard tasks go to paid models.

#llm#экономика#архитектура

→

2026-07-10

Context by patterns: one fact, one place, five layers

A system stays whole as long as every fact lives in one place and the logic sits on separate layers. I show it on my own working system: an 87-line map, 64 connectors, a term defined in a single file, and a one-day client install.

#aios#контекст#архитектура

→

2026-07-04

Search is moving into AI answers: three kinds of sites that win

Nearly half of commercial queries in the Russian-speaking web already come with an AI answer, and informational traffic is shrinking. In 2026 the winners are brands, primary sources and landing pages that close intent completely. I check every point against my own site: what's built, what's honestly unfinished.

#geo#seo#aios

→

2026-07-04

The Fable 5 lesson: one government order, and the model is gone

On June 12 the US Commerce Department ordered access to Anthropic's flagship model suspended. It came back on July 1. Nineteen days offline, a stress test for anyone who built automation on top of a single API.

#aios#модели#риски

→

2026-06-20

Your CRM is empty because the deals live in Telegram

Across the CIS, sales happen in messengers while the CRM gets filled in from memory at the end of the day, if at all. An AI layer reads the conversations itself: deal cards, statuses, reply drafts. What's actually automatable today in the CRM + Telegram + WhatsApp stack.

#aios#crm#мессенджеры

→

2026-06-16

While I sleep: the system catches every incoming

An email at 3am, a document dropped into a chat on Sunday, big project news landing on a holiday – a person misses these, the system doesn't. It's always online, catches every incoming signal, and decides whether to wake me or let it wait until morning.

#aios#автономия#входящие

→

2026-06-14

Reconciled to the cent: what real reconciliation means

Reconciling statements to the cent means understanding what each line actually does, not just adding up totals. How the system does it and what hides behind the word 'deposit'.

#финансы#автоматизация#данные

→

2026-06-14

Autonomy hours: the metric I build the system around

How many hours a day the business runs without me at the screen. How I measure it, what runs while I'm away, and why the target is four hours by December.

#aios#автономия#метрики

→

2026-06-14

Memory matters more than the model

A chatbot answers and forgets the conversation. The system remembers your business: counterparties, contracts, the history of decisions. The model swaps out in a day, the accumulated memory doesn't. That's where the moat runs.

#aios#контекст#архитектура

→

2026-06-14

Decisions from your phone: running the system with one tap

My system's interface is a Telegram chat, not a dashboard. Each morning a short list of what's stuck on my decision shows up, and I clear every item with one tap, without opening the laptop.

#aios#telegram#интерфейс

→

2026-06-14

A system live in one day

AI rollout usually means a project that runs for months. For the first client the whole system landed on the manager's laptop in a day, with a Telegram bot as the way in. Why it's that fast, and where the real work is.

#aios#внедрение#клиенты

→

2026-06-13

What local models can actually handle

Can a business run its operations – mail, bank, contracts – on a model sitting on your own hardware? Numbers from a live system: a local model covers 92% of the work, only a trimmed slice goes out with a log, and the line keeps moving toward local.

#безопасность#данные#архитектура

→

2026-05-30

What an AIOS is and why a business needs one

An AI operating system sorts your inbox at six in the morning while you sleep. How it differs from a chatbot, the five layers it's built from, and why a rollout starts with the context layer while the model plugs in last.

#aios#архитектура

→