Roman H.

Most n8n builds break the day something upstream changes. A model returns text in the wrong shape, a scrape comes back empty, a provider rate-limits at 2am, and the whole run either stalls or quietly ships garbage that nobody catches until it is in front of a customer. I build n8n automations with AI agents wired into them, and I run more than one model on purpose. When a provider is slow, wrong or expensive for the job, the workflow uses a cheap one to filter and the strong one only where it earns its cost. The graph expects things to fail and keeps going. WHAT I BUILD Multi-agent content engines. A thread or a page gets broken into structured JSON, then several drafting agents each run on their own model with their own voice rules. An AI judge scores every draft and kills any claim that is not on a list pulled live that same run. Nothing reaches a human queue unscored, and the AI physically cannot recommend a feature that does not exist. Scoring pipelines that spend money in the right order. A cheap model gates the volume, the expensive one only writes the full report on whatever survives. You see the reason behind every pass and every reject, so a filter that starts matching the wrong thing is visible instead of silent. Scraping and enrichment that does not trust one source. Apify and Firecrawl for the pull, structured parsers on the way out, and a fallback path for whatever comes back thin, so one bad page does not poison the whole batch. HOW IT HOLDS UP Every AI step is schema-enforced with an output parser, so a model that rambles does not break the node after it. Factual and product claims are bounded by data fetched on that run, not by whatever the model happens to remember. Model names live in one Config node, so swapping Gemini for Grok or OpenAI is a one-line edit, not a rewire of the graph. HOW I WORK We agree the spec before I build. Triggers, data, what happens when an API is down, what "done" means. You get a fixed price against that document instead of an open hourly. Every rejection is logged with its reason, so when something is wrong you can see why. You get the JSON, the credentials in your own accounts and a handover doc, so you are never locked to me. STACK n8n self-hosted and cloud. OpenRouter, OpenAI, Anthropic, Gemini, Grok. Apify and Firecrawl for scraping. Telegram, Slack, Reddit and LinkedIn sources. Supabase, Postgres, Airtable. Structured output parsers, AI judge agents, memory buffers, REST, webhooks, OAuth2, JavaScript. START HERE Tell me the workflow that keeps breaking, or the one you have not built yet because the AI part looks fragile. I will tell you what is worth automating, what is not and roughly what it costs. That part is free.