Multi-marketplace e-commerce.

The problem

A multi-channel e-commerce seller present on Central Europe's largest marketplaces, with extensive purchasing, logistics and planning operations — previously run largely by hand, in spreadsheets and inboxes. Every new sales channel and every new country added another layer of manual work.

What I built

An automation platform on which more than 200 workflows have been built — partly by me, and partly by the client's team, which I prepared for the job:

• Marketplaces: 7+ channels, 6 countries — Allegro (PL and CZ), eMAG (RO and HU), Kaufland, Empik, the Baltic markets: daily buy-box monitoring, price archives, daily margin snapshots and automated repricing — always with "what changed / what needs attention" reports and a task for a human, never blind.
• Agentic product research — an AI agent that, given an EAN, finds the product's composition in trusted sources and returns the result with a source and a confidence level. Alongside it, a pipeline generating product names, usage instructions and marketplace descriptions straight into the PIM.
• AI in customer service — a gateway classifying every helpdesk ticket into 10 intents, rolled out in stages: first measure accuracy on live traffic, only then switch on automated actions.
• The company's operational memory — every recorded meeting lands in a searchable base with a privacy filter; an assistant answers questions like "what did we agree with X in March" with a link to the source.
• An agent that keeps projects on rhythm — in a company with no middle-management layer, an AI agent reviews the growth-project portfolio in Asana every morning: it publishes a management note (project assessment, pending decisions, owners, deadlines), updates statuses when an assessment changes, and selectively chases tasks without an owner or decisions without an answer — with a daily reminder cap, so it supports the rhythm instead of spamming. It writes statuses and comments, but deliberately never edits anyone's tasks.
• Code where low-code ends — custom microservices backing the automations: a multi-platform scraper that handles anti-bot protections, a document converter into AI-readable format (released as open source), an Excel-processing service that preserves formatting, and product matching across platforms.
• An internal automation team — the most durable part of the engagement: thanks to n8n-mcp (my tool, where an AI agent builds the workflow for the employee), people create automations themselves without a long platform learning curve, and regular office hours serve to review, unblock and raise the bar. Automation stopped being an external service — it became a company capability.
• A foundation for scale — the platform runs on the client's own infrastructure, in a queue architecture with parallel workers, with daily backups and monitoring: every failure automatically becomes a task for the right person, with full context.

Models from several providers run side by side in production — each picked per task, not out of vendor loyalty.

The outcome

An automated back office at the scale of multiple marketplaces and countries: entering a new market is configuration, not a new hire. Most importantly — the system belongs to the team, not the consultant: the people who know the work automate it themselves, while I look after the architecture, the harder pieces and the platform's growth.