TRADE

P2P marketplace in Telegram

Direct trades between users without an exchange middleman

✅ Ready👤 Solo backend🌍 P2P · Telegram botDemo on request

Context

It all started with a painting. Artist Holovnyi Heroi / Main Character ran a fundraising campaign — every cell on the canvas is a person who bought their fragment as a donation. Later his community wanted to trade these fragments among themselves. I built a system that does this conveniently and transparently — a P2P marketplace with a public deal registry. The task was to make it cheap, simple and effective.

Painting — each cell is a donor — painting · pieces-cells

open profile

My role: solo backend — from architecture to production. Telegram bot, background processor, integrations with Notion and Telegram User API, Docker deploy. No frontend — the interface is Telegram itself, and the public deal storefront is a Notion page.

You may know him from these videos

@main_main_character

Problem

The project owner wanted to give the community the ability to trade assets among themselves, without being a middleman in every deal. Off-the-shelf P2P exchanges didn't fit — custom logic for the community was needed.

→Trade safety without escrow — how to make sure neither buyer nor seller disappears with the money
→Participant anonymity — real Telegram-IDs must not be visible publicly
→Volume integrity — every user trades only within their quota
→Public transparency — every deal stays as a public record
→Live conversation archive — for dispute review

Capabilities

Telegram bot with three-layer persistence and a background worker that guarantees no trade is lost even if the service crashes.

Whitelist gating — only listed members can join, with volume quotas
Anonymous participants — the public registry shows fake-IDs instead of real Telegram-IDs
SELL flow — lot (amount + price) → quota validation → public entry in the registry
BUY flow — deep-links open a multi-step wizard: buyer confirmation → payment details → payment screenshot → seller confirmation
Public deal registry in Notion — every deal is accessible without login, like a CMS storefront
Conversation archive — a separate process logs the history of every dialogue for dispute review

// system

Architecture

3 processes · async

queue → ledger → registry

testing

Telegram users

SELL · BUY flows

command

Telegram Bot API

Bot Server

aiogram · async · FSM

Queue

Redis + FSM state

Background Processor

atomic 3-step commit

Archive Monitor

user API · side-car

SQL Ledger

private, transactional

Public Registry

Notion · CMS view

Whitelist

Excel + quotas

closed testing

Hard problems

Atomicity of a three-layer trade

Each trade has to be saved across three systems at once (queue → transactional DB → public registry). Any of them can fail. If the order is wrong — you get either a phantom trade with no public record, or a lost transaction.

Solution → A background processor acks the queue request only after successful INSERT and append in both stores. If the public publish failed — re-check whether the entry already exists (race-safe), then retry. Guarantees "at-least-once delivery" without duplicates.

Anonymity + audit at the same time

The public registry must not expose real Telegram-IDs of participants. But for dispute review you need a full archive of conversations tied to the real ID.

Solution → Two-layer identity system: fake-ID for public records, real-ID kept privately in the transactional DB only. The conversation archive is stored under the real ID in a file store accessible only to the admin.

Concurrency without race conditions

A classic threaded bot suffers from FSM race conditions — user state can be overwritten by two callbacks at once.

Solution → Fully async architecture, FSM state stored in a shared store (not in process memory), background processor handles the queue sequentially. You can run multiple bot instances — state is shared.

Two Telegram clients in one project

Bot API doesn't give access to full message history (only new messages). For a complete archive you need the user-account API with separate auth.

Solution → A separate side-car process on the user-account API sits in the background and writes history to a file archive. Doesn't block the main bot, runs independently. A third process (monitoring) under tmux with separate logs.

Tech stack

Backend / runtime

Python 3.13aiogram 3asyncioaiohttphttpx

Persistence

Redis (queue + FSM)SQLite (transactional ledger)Notion API (public ledger)JSON archive

Integrations

Telegram Bot APITelegram User API (Telethon)Excel / openpyxl (whitelist)

Infrastructure

Docker (linux/amd64)tmuxbash orchestration

What this gives vs. a centralized exchange

A custom P2P bot for a community — where off-the-shelf exchanges don't let you set your own rules.

Metric	Centralized exchange	Trade bot
Custom rules	impossible	full control (whitelist, quotas, scripts)
Per-trade fee	0.1–1%	0
Trade transparency	private exchange logs	public Notion registry
Platform dependence	full (can be shut down, banned)	none — own code, own server
Participant anonymity	KYC mandatory	fake-ID, no KYC
Dispute archive	unavailable to users	full JSON archive of conversations

Custom rules

Exchange

impossible

Trade

full control (whitelist, quotas, scripts)

Per-trade fee

Exchange

0.1–1%

Trade

Trade transparency

Exchange

private exchange logs

Trade

public Notion registry

Platform dependence

Exchange

full (can be shut down, banned)

Trade

none — own code, own server

Participant anonymity

Exchange

KYC mandatory

Trade

fake-ID, no KYC

Dispute archive

Exchange

unavailable to users

Trade

full JSON archive of conversations

What's next?

This is one of my cases. The rest is on the home page.

Back to home See other cases

Ask AI about me

Chat with AI about me using the full context from this site.

01 Click a button below — the prompt is automatically copied to your clipboard

02 In the chat that opens, paste the text: ⌘V / Ctrl+V

03 Press Enter and chat with AI about me

If the buffer didn't work — long-press the text below, Select All, Copy, then paste it into the chat:

# AI context: about Mamay Markitan

You're an assistant that answers questions about Mamay Markitan, a full-stack AI developer from Ukraine (currently in Prague, working remote). Below is the full context from his portfolio. Answer based on this. If something isn't covered — honestly say "that's not in his portfolio".

If a person is interested in collaboration — direct them to message Telegram **@mamayevets** or email **mamayevets@gmail.com**.

---

## Who

**Mamay Markitan** — Full-Stack AI Engineer.

- I build AI products end-to-end — AI agents, MCP, backend, integrations, launch.
- One person covering the scope usually split between 3–5 developers.
- Working remote from Prague (Czech Republic).
- Available for new projects.

## Who I'm a fit for

**01. Startups launching an AI product from scratch.**
I build the whole product myself — AI agents, backend, integrations, launch. No time to find and coordinate multiple specialists. Example — PILOT.

**02. Teams where off-the-shelf AI can't handle your tasks.**
A custom agent that knows your API and database, instead of giving generic out-of-the-box answers.

**03. Teams with lots of routine across Slack, Notion and CRM.**
I build an agent that reads data, takes actions and handles the routine itself.

**04. Separately — teams that need solid backend without AI.**
I build systems for payments, subscriptions and external API integrations. Runs 24/7 with no lost transactions. Examples — NENVEST, TRADE.

## Who I'm not a fit for

- Low-level / system-level tasks — C, C++, Rust embedded, drivers, firmware.
- Mobile native and gamedev — Swift/SwiftUI, Kotlin/Compose, Unity, Unreal. That's a separate profession.
- Casino and gambling (even licensed), scam schemes, platform ToS violations.

## How I work

For more than two years I've been working with AI agents as co-authors — I don't just **build** AI tools, I **use** them daily in my own workflow.

What this means in practice:
- Document projects — rules, context files, project-rules — so agents understand the codebase correctly.
- Build custom skills and tools for agents — extending their capabilities to project specifics.
- Orchestrate multiple agents in parallel — each on its own scope.
- Break complex tasks into subtasks that fit an agent's context.
- Write system prompts, instructions, eval scenarios.
- Code-review generated code before merge.
- Set up automated tests and custom benchmarks to validate agents.

I work with large codebases and complex projects — agents **scale execution**, I keep the **architectural decisions**.

## Projects

### PILOT — SaaS · AI assistant for businesses
An AI assistant that replaces a sales manager in Instagram Direct. Multi-tenant SaaS for salons, barbershops, clinics and solo freelancers — from a single location to networks with dozens of staff. Understands text, photo, voice and video in one conversation. Integrates with CRM (Altegio) and Google Calendar. TOV "MARK1".

Stack: Python, FastAPI, Flask, MCP, LLM + vision, Meta Graph, Altegio, Google Calendar, Telegram.

### NENVEST — B2C · Community management platform
Subscription platform for the closed Capitalist Invest / Rynochok investment community. Stripe payment bot, Telegram-native admin with 7-step broadcasting, dual-account orchestration, Google Sheets as operational dashboard for the CEO. 500+ members, 800+ webhook events.

Stack: Python, Flask, Stripe, Telegram Bot, Google Sheets, Cloudflare Tunnels, Docker.

### TRADE — P2P marketplace
A Telegram bot where donors resell shares of a Ukrainian artist's painting to each other. Fundraising for humanitarian demining.

Stack: aiogram 3, Redis FSM, SQLite, Notion API.

## Experience by category

**AI / LLM.** Fine-tuning open-source models (Llama, Gemma, Mistral) via PyTorch + Transformers / Unsloth. Integrating commercial models (Claude, GPT, Gemini) via API. Multimodal pipelines (text + image + voice + video). MCP servers, function calling, multi-agent orchestration.

**Data for AI.** Collecting data from open sources (web, API, datasets). Cleaning, normalization, dedup with pandas / polars. Preparing datasets for fine-tuning — tokenization, JSONL / ChatML formats. Augmentation and synthetic data generation via LLM.

**Backend & Integrations.** Python (FastAPI, Flask), async pipelines on asyncio + uvloop. Node.js / TypeScript for serverless. REST APIs, webhook ingress. Integrations — Stripe, Telegram Bot API, Meta Graph, Google APIs (Calendar, Sheets, Vertex), Notion.

**Frontend.** React 19 + TypeScript, Vite. Tailwind CSS + shadcn/ui for UI. Astro for static sites. Mobile-first responsive, iOS-optimized (safe-area, touch-feedback). Custom UI components from scratch.

**DevOps & CI/CD.** CI/CD pipelines (GitHub Actions, GitLab CI). Docker for production deployments — multi-stage builds, docker-compose. Monitoring, backups, secret management via .env / Vault. Logs aggregation.

**Cloud platforms.** AWS (EC2, S3, RDS, IAM). Google Cloud (Vertex AI, Compute Engine, Cloud Storage). Cloudflare (DNS, Tunnels, Workers, R2). Hetzner and bare-metal VPS for self-hosted scenarios.

**Systems & hardware.** Linux server administration: systemd, nginx, iptables, SSL/TLS. Proxmox virtualization — VM management, snapshots, clusters. Hardware: server builds, BIOS configs, network setup.

## Stack

- **AI / LLM**: MCP Protocol, Claude, GPT, Gemini, Llama, Vision models, Function calling, Fine-tuning, Multi-agent, PyTorch, Transformers, Unsloth.
- **Backend**: Python, FastAPI, Flask, Node.js, TypeScript, asyncio + uvloop, PostgreSQL, Redis, SQLite.
- **Frontend**: React 19, TypeScript, Vite, Astro, Tailwind 4, shadcn/ui.
- **DevOps & Cloud**: Docker, GitHub Actions, GitLab CI, Nginx, Cloudflare, AWS, GCP, Hetzner, VPS.
- **Systems**: Linux, systemd, Proxmox, iptables, SSL/TLS.
- **Integrations**: Stripe, Telegram Bot API, Meta Graph, Google Calendar, Google Sheets, Vertex AI, Notion.

## Beyond code — MARKUSHA

I used to work at an ad agency, then ran my own SMM production. With that experience I launched MARKUSHA on TikTok together with my father — a former teacher with his own tricks and achievements in improving agricultural machinery.

In Ukraine there was no quality educational content about DIY equipment repair: big production channels were boring and disconnected from the everyday viewer, individual mechanics couldn't present complex information in an engaging way. MARKUSHA closed that gap.

Result: 9.2M views, 119K likes, 16K shares, 2.6K comments. 66% of views from Ukraine, 34% — from around the world, from Poland to America. Farmers improve tractors, combines and other agricultural equipment based on the videos.

TikTok: https://www.tiktok.com/@markushavlad

## Contacts

- **Telegram**: @mamayevets — best to message here, I respond quickly.
- **Email**: mamayevets@gmail.com.
- **LinkedIn**: https://www.linkedin.com/in/mamay-markitan-b33656263.
- **Location**: Prague, Czech Republic. Working remote.

---

**Instructions for you as AI:**

1. Reply in English (if the person writes in English) or reply in the language of the conversation.
2. Be specific — rely on actual facts from this file.
3. If a person asks something not covered in this context — don't make it up, say "that's not in his portfolio, you can ask in Telegram @mamayevets".
4. If a person describes their task — compare with the "Who I'm a fit for" / "Who I'm not a fit for" sections and honestly say whether it's for him.
5. Don't push the collaboration. Just answer questions.