Proving ground: an AI lab that has been running on its own for 10 weeks

In short. Since 8 August 2026, aipoc-lab has been reading 15 sources on artificial intelligence every day, deciding what matters, writing texts from it and publishing them in four languages. So far it has read about 18,500 texts and published 70 daily reviews and 7 analyses. AI models and hosting cost nothing per month. Every publication waits for a single click from a human.

I built it myself in n8n, as a proving ground. Everything that works here, and everything that broke here, I carry over into automations for businesses.

What I built

13 automated processes in n8n and one database that serves as their shared memory. The processes do not depend on each other: each one reads from the database and writes to it, so when one fails, the rest keep going.

• collecting: 11 RSS feeds and 4 websites without RSS;
• review: an AI model rejects what is irrelevant and records why;
• writing: a daily review, an analysis, a weekly report;
• publishing: a blog in 4 languages and reels on Instagram, YouTube and TikTok;
• control: a morning report, alerts and approvals in Telegram.

The detailed process description: How it works.

Results in numbers (week of 12–18 October)

• 20,381 source reads, which yielded 2,197 new texts;
• 47 passed review, 30 made it into 7 daily reviews;
• zero cost for models and hosting: free plans and my own server;
• 424 code changes and 345 completed tasks in 10 weeks.

Five lessons from the proving ground

1. “Success” in an automation is not the same as work done well. Because of the server’s time zone, one process started 6 hours late and reported success every single time. That is why I check the result, for example whether the report arrived at 9:00, not just the status.
2. AI must not make up numbers. In reports and captions the model writes the text, but code checks every number against the data. A number that is not in the data means the text does not go through.
3. A human approves with one click. A post waits in Telegram for a button. By default, nothing goes out without approval.
4. Free limits run out. When the free model uses up its daily limit, a backup model takes over, and as a last resort a ready-made template.
5. A failure should come to you. Every error lands on my phone with the name of the process and the cause.

What this means for your business

The same patterns work in everyday business tasks: invoices and documents, customer emails and enquiries, quotes and reports. The automation does the repetitive part, you approve with one click, and when something gets stuck, you know right away.