I design and build backend systems, AI-powered workflows, automation platforms and business software with Python.
I build systems.
Code is the last step. Before it comes the problem, the people who live with it, and the shape of a system that makes it go away.
From APIs and databases to AI workflows and automation, I focus on turning complex business requirements into reliable software systems.
→ Django / FastAPI
APIs, domain logic, auth, data access — the part that has to be right.
→ LLM / AI workflows
Tool calling and agents placed inside real processes, with real limits.
→ Celery / Redis / integrations
Background jobs, schedules, webhooks and bots that remove manual work.
→ PostgreSQL / APIs / distributed workflows
Schemas, boundaries and reliability — designed to survive change.
A Telegram outreach CRM driven by an LLM tool-calling agent. It finds and ranks channels, discovers their admins, drafts outreach and runs campaigns — and every high-risk action waits for a human to approve it.
A dispatch platform that runs inside Telegram. Orders come in, drivers are matched, routes and statuses are tracked — with nothing new to install.
A service management platform for a landscaping business: requests, quotes, crew scheduling and job tracking in one system.
A productivity ecosystem where chat, tasks, notes and voice are connected through a shared AI layer that can act — not just answer.
A Telegram outreach CRM driven by an LLM tool-calling agent. It finds and ranks channels, discovers their admins, drafts outreach and runs campaigns — and every high-risk action waits for a human to approve it. Below is the system as it runs: seven stages, each one removing a piece of manual work.
The agent turns an operator's request into a channel search. The Telethon worker queries Telegram's public search and scores the results for relevance.
New channels are sent to the LLM in batches as a Celery job. A structured-output schema returns a quality score and category, and flags duplicates.
Channels are joined through the operator's own Telegram session. Outbound actions are held to a daily cap, working hours and randomised pacing, and freeze during Telegram flood penalties.
Admin lists are fetched through the Telegram API, de-duplicated across channels and saved as contacts.
Leads move through a status pipeline — from admin found to message sent, replied, negotiation and follow-up — with a full status history.
The LLM drafts each message. Sending, launching a campaign or deleting data pauses the agent; the operator approves or rejects in Telegram, and the database only releases approved drafts.
Approved messages go out through the worker. Replies are classified and summarised by the LLM, follow-ups are scheduled, and a watchdog flags campaigns that stall.
Manual Telegram outreach was slow and hard to manage. Finding relevant channels, working out who ran them, remembering who had already been contacted and writing every message by hand did not scale past a handful of prospects. Handing that work to an AI raised the opposite problem: nothing should reach a real person without someone approving it.
An operator writes requests to a Telegram control bot. An OpenAI function-calling agent chooses from 34 tools — search, rank, join, fetch admins, save contacts, draft, send, organise folders, launch campaigns — runs them through a Telethon worker and reads each result before deciding the next step. Campaigns run as a database-backed state machine on Celery. High-risk actions pause the agent for approval, and a first message can only be sent once a human has approved the draft.
Deployed on a VPS and used in a pilot, the system discovered 52 channels and 82 admins and routed 366 outreach drafts through the approval gate: 17 were sent after a human approved them, 10 were rejected. The agent made 90 traced tool calls. Search, discovery, drafting and sending now run as one workflow instead of ad-hoc chats.
Chosen for what they do in production, not for a logo wall. Hover anything for the role it plays.
The core: domain logic, APIs, the code that has to be right.
5 in daily use
Schemas that survive change and queries that stay fast.
4 in daily use
Everything that runs when nobody is looking.
5 in daily use
Language models inside real workflows, with real boundaries.
4 in daily use
Reproducible, observable, boring in the best way.
5 in daily use
Products that live where the users already are.
3 in daily use
Every project runs through the same five steps. The order matters more than the tools — most expensive mistakes happen before the first line of code.
Before any code: who uses it, what breaks today, and what 'better' looks like in numbers. Most failed systems solved the wrong problem well.
Data model, boundaries, integrations. Decide what is synchronous, what becomes a background job, and where the risk lives.
Business logic first — typed, tested, explicit. Then the API that exposes it. The interface can change; the core has to hold.
Background workers, integrations and AI steps that remove the manual work around the core. Software should do the repetitive part.
Logs, metrics and feedback loops. Systems get better when you can see them — and worse when you can't.
IT / Software Projects · 2026 — Present
Designing and building backend systems, AI-assisted workflows and automation platforms for business clients.
Shipped an AI-assisted outreach CRM and a Telegram-native logistics platform.
01 / 04Independent products & client work · 2025
Owned products end to end: requirements, data model, APIs, bots, deployment.
Turned recurring business problems into working, maintained software.
02 / 04Freelance & startup projects · 2024
Django / DRF APIs, PostgreSQL schemas, third-party integrations and background processing.
Delivered production APIs for early-stage products under real deadlines.
03 / 04Python · Databases · Systems · 2023
Deep focus on fundamentals: Python, SQL, HTTP, Linux, and how systems fail.
Built the base every project since has stood on.
04 / 04Let's build it.
Tell me what is slow, manual or fragile in your business. I'll tell you what a system for it would look like — and what it would take to build.