I build AI assistants 
that run your routine.

For one person, or for one process in your company.

For yourself, or for your company

For yourself

One assistant that knows your life.

For founders, engineers and anyone whose own admin never gets done.

  • Your health data explained every evening, not charted
  • Your inbox, calendar and tasks triaged before you wake up
  • Everything you've ever told it, searchable months later
From $900 for the audit. See personal pricing

For your company

One process, automated end to end.

For teams of 5–50 in e-commerce, services, clinics and support.

  • Customer reviews and questions answered in your tone, risky ones escalated with a draft
  • A daily report in words, from the numbers you already collect
  • Invoices and documents from a photo into structured rows
From $2,500 for the audit. See business pricing

A solid mark already runs in my own setup. A hollow one is a build.

For yourself

  • Food and medication from a photo — the dish identified, calories and macros logged. (Live)
  • Chats and channels: condensed to the three things that actually matter. (Live)
  • Documents and deadlines: visas, insurance, renewals, chased before they bite.
  • A personal CRM: who is waiting on you, who you promised what, whose birthday is next week.
  • Money: subscription audit, spending digest from statements, bills before they are late.

For your company

  • Team digest: support channels and chats, deduplicated against what was already reported. (Live)
  • Screening: spam, cold pitches and real leads sorted before a person looks. (Live)
  • Company memory: decisions, promises, who said what and when, searchable months later. (Live)
  • Marketplace sellers: Wildberries and Ozon reviews, stock alerts, a morning report per store.
  • Monitoring: competitor prices, regulation changes in your niche, news that actually concerns you.

What exists when it's finished

  • A working agent on your machine or in your cloud account, using your own accounts and keys
  • A written rule set: its tone, what it decides alone, what it always escalates
  • A log of every action it takes, so you can audit what it said and did
  • A handover document detailed enough for another engineer to take over
  • The code and the repository, yours, no licence and no lock-in

Four steps, four to six weeks

  1. Call, 30 minutes.You describe the work you keep postponing, and I say plainly whether an agent helps.
  2. Audit and plan, one week.What can be automated, what can't, the architecture, the integrations and an honest monthly running cost.
  3. Build, three to five weeks.A working agent on your infrastructure, your data, your tone, escalating to your people.
  4. Care, monthly.I watch it, fix drift and add skills as you find new work for it. Cancel any month.

The audit is a document you keep, even if you stop there and never commission the build.

Who you're dealing with

Vlad Arbatov, at night against city lights

I'm a founding engineer at a privacy-first AI company. Before that I was VP of Engineering in streaming and telemetry, and built mapping infrastructure at Mapbox. I run one of these agents — Volna — on my own hardware every day.

Proof

Volna

Volna's portrait — a stylized illustration of a young woman with short dark hair

Volna is the assistant I built for myself, and she has run on my own server every day since 27 February 2026 without being rebuilt. She is the reason I am willing to sell this work: every technique on this page is one I maintain daily, not one I read about.

What it looks like

Personal data blurred in the source files.
Screenshot pending
Dashboard
Screenshot pending
Evening health story
Screenshot pending
Food from a photo
Screenshot pending
Nightly skill edit

Measured on 25 July 2026

  • 5 monthsrunning since 27 February 2026, never rebuilt from scratch
  • 25scheduled jobs
  • 32skills
  • 11connected services
134
code changes across 54 files, ~13,800 lines
several hundred
available tools
1,170
runs in the last 32 days of retained transcripts
4,187
tool calls in the same 32 days
1,881
replies in the same 32 days
95
food log entries
129
medication log entries
4 of 5
processes up nine days since the last deploy

Where the time goes

My estimate, method below.
WorkManual, per day
Health review, morning and evening15 min
Chat and channel digest25 min
Tasks, calendar, subscriptions8 min
Post drafts and replies15 min
News in my own topics10 min
Meeting notes5 min
One-off lookups, notes, food photos, “what did I say back then”20 min

≈ 1 h 40 min a day 10–12 hours a week roughly 200 hours over five months

What that number is really made of

Half of this is work I would have done by hand. The other half would never have happened at all. A first build lands nearer 3–5 hours a week and grows as more of your work moves in.

Why it's still open in month six

Most agent projects stop being opened by week three, and the model is rarely the reason.

Memory that survives compaction

A maintained portrait, daily notes and an entity graph are rebuilt outside the conversation and injected into every run, so nothing depends on chat history.

Silence as a first-class answer

Scheduled jobs are worthless if they generate noise, so saying nothing is a valid result — and it fires more often than a message does.

Spend ceilings

Model tiers per task type, a cheap router in front and a hard cap: an always-on agent without a ceiling is an open tap.

Escalation, and self-correction

Anything ambiguous, angry or expensive goes to a person with a draft attached, and every night the agent reads its own transcripts and rewrites the instructions it got wrong.

Prices in USD

What it costs

What it costsFor yourselfFor your company
Audit and plan1 weekfrom $900from $2,500
Build3–5 weeks$4,000–9,000$12,000–25,000
Caremonthly$400–800 / mo$1,500–3,000 / mo

Model usage is billed to your own account, and I set the ceiling with you before the build starts.

What's in the build

What I don't promise

  • Full autonomy.The agent makes mistakes. Escalation to a person is part of the build, not an afterthought.
  • A flat monthly bill without a ceiling.Model spend is real and has to be capped deliberately.
  • Instant setup.Most of the calendar goes to access, accounts and writing down how your team actually works — not to code.
  • Smooth access to frontier models from inside Russia.It's a separate problem and we settle it before signing, not after.

Who this isn't for

  • Anyone who wants a product they can sign up for today.
  • Anyone who wants an agent that acts with no human review.
  • Anyone who can't give access to the systems the work actually lives in.

Before you book

Where does my data live?
On your machine or your cloud account. Nothing is required to pass through mine. For Russian clients, hosting inside Russia is the default.
What if you disappear?
You own the code and the repository outright, and everything is plain files with the processes written down. Another engineer can pick it up — that's a build requirement, not a hope.
What does it cost to run per month?
It depends on volume. I model it during the audit week and set a hard ceiling with you before anything is built.
Will it answer customers as us?
In your tone, by rules you approve, and with anything risky routed to a person first — never as an anonymous bot.
How do we start small?
One process, four weeks, fixed price. If it doesn't earn its keep, we stop there.

Bring me the thing you keep postponing.

Thirty minutes, no deck. If an agent is the wrong answer, I'll say so on the call.