AI developer · building with Claude Code
I build AI agents, assistants and automations that take routine work off your team.
A bot answers your clients at night and books them in. An agent sorts the inbox and moves data between your tools. I check every piece before launch, and you can test a bot right here.
For employers
3+ years with AI tools (AI content since Aug 2023); building software with Claude Code since Jun 2025.
Multi-agent LLM pipelines are the core of my work: I have built 57 working agent-orchestration workflows, with independent checkers and my final approval. I connect services through APIs and webhooks and cover the logic with automated tests.
CV on request: Telegram @shov69 · sherlock753cc@gmail.com
What changes for your business
Most teams lose hours every day to work that repeats: the same answers, the same copy-paste, the same reports. I hand that work to bots and AI agents, so your staff spend their time on the work only people can do, and you can grow without hiring extra hands for routine.
Clients write at night and on weekends, and nobody answers
A bot answers from your prices and FAQ, books a free slot and sends you the lead with Confirm and Decline buttons.
Test the bot
Your team answers the same questions all day
An AI assistant answers from your own documents and shows the source. When the documents do not say, it says so and passes the question to a person.
See the assistants
Someone copies data between tools by hand
A script or an agent collects the data, moves it between your tools and sends the report to Telegram or a spreadsheet.
See the automations
Requests pile up in email, forms and chats
An agent reads every message, pulls out the name, contact, budget and urgency, drafts a reply and sends the hot ones to you first.
See the inbox agentWork
Demos run on fictional data; the bot you can test right on this page. The code is private; I walk through it and show the demos on a screen-share call.
AI assistants and agents
An assistant answers clients from your own documents and shows the source. An agent reads email and requests, pulls out what matters, drafts a reply and hands you the hot leads. When there is no answer, the assistant says so.
Answers site visitors and emails in the voice of the business, keeps every conversation, and pings me in Telegram, where I read it, take over or cancel an answer. Before launch a second check in a separate session found 13 issues; all were fixed with regression tests.
≈ 3 cents of API cost per answer, measured on the live setup
Ask about a fictional company: the answer comes with numbered sources, and next to it you see which parts of the documents the search found. A question outside the documents gets an honest “not in the documents”.
The agent sorts each message, pulls out the name, contact, budget and urgency, drafts a reply and sends hot leads to the owner at once. You see what the AI did and what the rules did.
Downloads Telegram channels, transcribes voice and video on the GPU, and answers questions with a link to the exact source and timestamp. Plus full-text search over 173,000 messages.
4,075 chunks in the vector index from 29 sources
Can build Also on request:
Need an assistant on your own documents? Tell me what it should know and whom it answers.
Routine work automated
Scripts and agents that collect data, connect your tools and send reports. Flows people often build in n8n I build in code, so they are easy to change and covered by tests.
Crawls a test shop’s catalogue on its own, stores prices, and shows the changes, a chart and a CSV export. It refreshes itself.
Finds videos that beat their channel’s norm, collects public comments and transcripts with no API key and shows it all in filterable dashboards.
315,000 comments, 1,473 transcripts and 867 videos collected
My own multi-page site: a cost calculator, an interactive map, a short quiz, an AI chat and a contact form that pings me in Telegram.
31 areas on the interactive map, 20 pages
Five questions, an instant estimate, the lead saved and a card for the owner in Telegram. A fictional renovation company.
A commercial proposal as one web page with evidence pop-ups and variants set by the link; seven deploys with automated browser checks, 13 client edits shipped in one round.
Turns an approved script into a finished 16-minute video with AI voice, footage from a 2,700-clip library and real chapters; 19 rendered in one week.
Drafts titles, chapters, tags and a pinned comment from a video summary, and finds videos that outperformed their channel.
Can build Also on request:
Is someone doing a job by hand every day? Describe it and I’ll suggest how to automate it.
Bots
A Telegram bot answers questions, books visits and hands each request to you with buttons. A Mini App inside Telegram shows the team every lead, its status and the stats.
A fictional print shop in a chat-app window. The customer writes first and the bot answers as the shop: it prices business cards, flyers, posters, banners, stickers and brochures with every line of the math shown, takes the order step by step and shows the card the owner would get, with all the details. Prices come from the price list, not from the AI; the AI only helps understand free text. Test it right here; it speaks English and Russian.
28 of 28 checks pass in its live end-to-end test
A fictional studio, so you can test it like a client. It answers from the FAQ with AI, books in four taps, sends the owner a card with Confirm and Decline buttons, and has an owner view. If the AI is down, it answers from the FAQ by keywords and never goes silent. Test it in Telegram; it speaks English and Russian.
24 of 24 checks pass in its end-to-end test
A Telegram bot qualifies people with a short form and hands each lead to a consultant. The team works the leads in a Telegram mini-app: filters, statuses, owners, comments, reminders and funnel stats.
Live since September 2026 on Cloudflare Workers, 17 database migrations
Leads, statuses, reminders, stats and a morning digest. Every visitor gets a private sandbox with test data; in a browser it opens the same way as in Telegram.
Messages from a site chat, email, a contact form and YouTube comments land in one Telegram Mini App, where I reply, snooze, add notes, or take a conversation over from the AI and hand it back.
208 of 208 automated tests pass
Collects similar listings every day, tracks the object’s price and views, and sends the seller a report every evening. Tied to the agent’s chat, so I show it on a call.
Can build Also on request:
Need a bot for your task? Describe it in a couple of lines; I reply with questions and a first step.
How I work
An AI agent in Claude Code writes the code, and I answer for the result. So every step has a check.
1You describe what you need. I answer with questions about the logic and how we will know it is done.
2I write down what we build, step by step, and how each part will be checked.
3Claude Code writes it from the plan; I steer it and read every change.
4The core logic gets tests that run before every launch.
5A separate session that did not write the code looks for bugs in it. Every finding is fixed and covered by a test.
6I check it on a test link or in a test chat myself, then you test it the same way before it goes live.
Example: on the AI sales assistant the second check found 13 issues before launch. All were fixed, and 172 of 172 tests pass.
Contact
Describe the task in a few lines. I reply with questions and a small first step.
