Artem Samsonov

Your AI works.Now put it in front of customers.

I take one process — quotes, support, or search over your own documents — and build everything it needs to go live with customers: logins, access control, limits, per-customer usage tracking, payments, logs and alerts.

Fixed price, fixed timeline, the code stays in your repo.

Not another chatbot. Not an n8n workflow. A customer-facing service, ready to run.

Book 20 minutesEmail me

Sound familiar?

  • It works in the demo, but putting it in front of customers is scary.
  • Login works every other time, and I can't say who sees whose data.
  • Payments are bolted on, but subscriptions have a mind of their own.
  • Fix one thing and another breaks — and there's no telling what the last prompt did.
  • The token bill is growing faster than the user base, and nobody knows why.
  • Half the work is still done by hand, and it's unclear where to start automating.

What I offer

Diagnostic: what's keeping your AI from customers

up to €3,500up to 5 working days

A founder with a stalled pilot, a team about to roll an AI feature out to everyone, or the owner of an app built in Lovable, Replit or Cursor.

  • I review your code, access and infrastructure.
  • I list the problems by severity, each with the risk explained and an estimate of how long the fix takes. You can hand this list to any other contractor.
  • I fix the most critical one right away.
  • If everything checks out, I'll say so plainly.
Book 20 minutes

Your process with AI

up to €10,500up to 3 weeks

Something is still done by hand that should have been automated long ago: quotes, support, search over documents, or invoice processing.

  • We pick one process and put in writing what will count as the result.
  • I automate it and connect it to the systems it needs.
  • Anything the system is about to change gets confirmed by a human first.
  • I set up logins and permissions: who sees what and who can change what.
  • I add limits, logs and monitoring, and test it under load.
  • I hand over the documentation and walk your developer through it on a call.
Book 20 minutes

Your product inside Claude and ChatGPT

up to €10,500up to 3 weeks

Users ask: “Does this work from Claude or ChatGPT?” I'll make it work.

  • I build a tool set around the agent's real tasks instead of just exposing every API method.
  • I set up sign-in and permissions: each customer sees only their own data.
  • I make errors clear enough that the model stops repeating the same call in a loop.
  • I add cost tracking per customer.
  • I set up limits, logs and monitoring, and test it under load.
  • If needed, I publish the product in the Claude and ChatGPT directories.
Book 20 minutes

Ongoing support

€700–3,500 / momonthly

  • Watch — €700/mo, about one working day a month: incident triage, control of what requests cost, dependency and security updates.
  • A monthly report: what worked, what broke, what it cost.
  • Fractional engineer — €3,500/mo, about a day a week. All of the above plus a task queue, review of code your AI writes, and a seat in planning.
Book 20 minutes

If your task is smaller than a package or doesn't look like any of them, book the call — I'll quote a timeline and a price for it.

Eight questions to ask before customers touch your AI

  1. Can one customer see another customer's data?
  2. Are any keys or secrets sitting out in the open?
  3. How many customers can the service handle?
  4. What happens when an external service errors out or hits a rate limit?
  5. Is there a spend ceiling, and do you know the cost per user per month?
  6. Are dev and prod actually separate?
  7. Can you roll back to yesterday in five minutes?
  8. Who finds out first when it breaks at night?

What happens when the answers are missing

  • 3.7M records — chats, call recordings and their transcripts with customer details from 2024–2026 — sat in three open databases behind Sears Home Services' support chatbot until an outside researcher found them.Cybernews, March 2026
  • Lovable: from February 3 to April 20, 2026, any user of the platform with a link could open the code and chat history of public projects. The first researcher report came on February 22; the fix shipped on April 20, after public disclosure.Lovable's postmortem, April 2026
  • A Cursor agent on Claude Opus 4.6 was working in staging, hit a credentials error, found a full-access Railway token in an unrelated file and deleted PocketOS's production database together with its backups in one request. Nine seconds.PocketOS founder's post, April 2026
  • An agent with AWS keys and a deadline spun up five 48-core servers to scan the DN42 network and ran up $6,531 over one weekend. The operator stopped it only when the charges showed up.Lan Tian's account, DN42, May 2026
  • 28.65M secrets leaked into public GitHub repos in 2025; commits made with an AI assistant leak secrets twice as often; 24,008 secrets sat in MCP config files.GitGuardian, March 2026
  • 44% of code-generation tasks introduce a vulnerability: the secure-code rate is 56%, same as in the first report, and bigger or coding-specialised models do no better.Veracode, 100+ models, July 2026
Book 20 minutes

What I built

ComedyCat

2025

This could be search over your own documents.

  • 1,500 shows, 70,000 jokes, 680 comedians
  • recommendations from what you like
  • search by meaning, not by keywords

Description & Goals

A social network for standup fans. Your own rating of shows, what your friends have watched, recommendations. Type “what an immigrant struggles with” and you get a selection by meaning, not by matching words.

Inside is what I build for clients in “Your process with AI”: the catalog refreshes automatically, jokes get extracted, indexed, and surface in search and recommendations.

Next.jsSupabasePostgreSQLpgvectorn8n
Open

Transcriptor MCP

2025open source

A service ComedyCat pulls its transcripts through every day. You can see the whole kitchen: code, images, CI, metrics, load tests.

  • 8 tools for 11 video platforms
  • Docker images, CI publishing, Redis cache
  • Prometheus metrics, k6 load tests

Description & Goals

Subtitles, metadata, chapters, playlists, frames and search — everything your agent needs to “read” public videos.

Listed in the ChatGPT app directory. This could be “Your product inside Claude and ChatGPT.”

TypeScriptMCPRedisWhisperDocker
OpenSource In ChatGPT

MCPal

2026

Sign-in, limits, usage metering and payments for any MCP server. The layer between “works for me” and “works for customers”.

  • sign-in from Claude, ChatGPT and Cursor
  • usage metering and a limit per tool
  • subscription and payments on top

Description & Goals

A developer plugs in their MCP server; users of Claude, ChatGPT and Cursor sign in via OAuth, use a free quota and subscribe when they want more. Different servers and their users in one system, and nobody sees anyone else's data. It's the same layer I build around clients' AI; here it runs in production as a product of its own.

Node.jsFastifyPostgreSQLOAuth
Open

How I'm different

  1. I build prototypes with the same tools, in the same evening. The difference starts the next day — when the first hundred users and the first payment arrive and someone has to be accountable.
  2. I don't propose a rewrite. I read AI-generated code and understand what it was trying to do.
  3. Fixed price, a success metric signed before the first line of code, code in your repo. You can walk after any stage. I hand over finished work with documentation and a roadmap.
  4. For five years in fintech I answered for the product, not just my slice of the code: conversion, tests, releases, monitoring, a team of ten engineers.
  5. I'll tell you before the contract if there's nothing worth automating or the economics don't add up.

About me

I grew from frontend engineer to R&D lead at Russia's largest financial marketplace. For the last five years I ran cross-functional teams of up to ten engineers, on both the organisational and the technical side. I was responsible for the product, system design, quality, security, keeping production running without interruption, monitoring and alerting. That's exactly the layer I now build around AI. A physics-and-maths foundation from MEPhI helps here. Since late 2025 I've been on my own, building my own products end to end. All of them are designed, built and maintained by me personally. I live in Barcelona, and work in Russian and English.

Not for me

  • Training or fine-tuning models.
  • Design work.
  • One-off workflows with no one operating them afterwards.
  • Work for legal entities registered in Russia or Belarus (EU sanctions).