Keeping an online store catalog in sync with n8n + a local LLM (Ollama)

I run online stores as a side project, and the catalog was always the part that quietly broke: supplier prices and stock change constantly, so the listings drift out of date until a customer orders something that’s actually unavailable or mispriced. I wanted to remove the manual checking entirely and only spend human attention on the cases that really need it.

So I built a small automation that keeps the catalog in sync on its own. Here’s what it does, the stack, and — more importantly — what I learned along the way.

What it does

• Pulls catalog, stock, and price data from the supplier source
• Normalizes messy product fields and matches them to the store’s listings
• Aligns stock and price automatically when there’s a confident match
• Flags exceptions for review — unmatched SKUs, suspicious price jumps, items going out of stock
• Runs on a schedule, so the catalog stays current without anyone watching it

The stack

• n8n — orchestration and scheduled workflows
• Ollama (local LLMs) — cleaning/normalizing unstructured product fields, cost-free and offline
• Python — data handling and the deterministic checks around the LLM
• MCP connectors — to wire the tools together
• Environment variables — all credentials stay in .env, nothing secret in the code

How it works (high level)

1. A scheduled n8n workflow triggers the run.
2. Supplier data is fetched and passed through a cleaning step — the LLM where the data is messy, plain rules where it isn’t.
3. Items are matched to existing store listings.
4. Confident matches get their stock/price updated automatically; everything uncertain goes to an exceptions list for a quick human pass.

What I learned

• Orchestrating reliable scheduled jobs in n8n instead of one-off scripts.
• Running LLMs locally with Ollama — what they’re genuinely good at (cleaning fuzzy data) and where they’re not.
• When not to trust the model. I also do paid AI-output evaluation, and that habit shaped this build: anything that touches price or stock goes through deterministic checks, not the LLM alone.
• Designing for the exception path, not just the happy path.