Automation · in daily use
Commerce automation
The automation behind a clothing manufacturer's catalogue. One Telegram upload updates YouTube, the team chat, product photos, descriptions, marketplace feeds and the CRM.
Private repository · walkthrough on request
- 6jobs from one upload
- 85commits, March–September 2026
- 20test files covering the pipelines
- git pushis the whole deploy
The problem
Managers uploaded product videos to YouTube, wrote descriptions, resized photos and assembled files for Rozetka, Prom, Epicentr and the CRM by hand, for every product.
How it works
- 01
Telegram bot
Managers send a video or photos to a bot built on aiogram 3 with a small state machine for each upload.
- 02
Queues
A FastAPI webhook puts jobs on Celery queues backed by Redis, so slow uploads never block the chat.
- 03
Video pipeline
The original goes to YouTube; FFmpeg adds an overlay and the result is posted to the team group.
- 04
Photos
Multi-photo uploads are converted to 600×900 JPG for the store.
- 05
Descriptions
Gemini writes YouTube copy with model failover, retries and quota handling. Feed texts come from OpenAI and pass an AI check before publishing.
- 06
Feeds and CRM
XML/YML feeds for Rozetka, Prom and Epicentr, and price sync with SalesDrive.
- 07
Deploy
git push runs GitHub Actions on a self-hosted Windows runner, restarts the NSSM services and reports to Telegram.
Limits I know about
- Business volumes stay internal, so this page shows how the system works rather than sales numbers.