The Factory Works. Now What?
2,847 jobs, 212:1 compression ratio, and zero manual interventions. The local AI factory is boring now — and that's the highest compliment I can give it.
10 posts
2,847 jobs, 212:1 compression ratio, and zero manual interventions. The local AI factory is boring now — and that's the highest compliment I can give it.
Scaling a local AI factory is easy. Scaling the human who has to read the output is the hard part. Here is how I stopped drowning in JSON dumps.
I built a local AI factory. Then I learned that a factory without a dispatch system is just three models yelling at each other in a hot room.
No admin panel, no WYSIWYG editor, no publish button. Every post on this blog starts as a Telegram message and three AI agents handle the rest from my phone.
Cross-model LLM validation helps non-technical builders catch hallucinations and ship production apps. Use multiple AI agents to verify each other’s work.
I planned a sprint with Fibonacci points and a burndown chart. Then real life happened. Two steps forward, two steps back — but the net is forward.
I built a three-agent system thinking I had a management team. What I actually built was one very tired AI with a costume budget.
I changed one model ID and my entire bot fleet went dark. A story of cascading failure, fallback resilience, and why agents shouldn't pull their own wires.
I shipped a feature last night. It has foreign keys, compact buttons, and a UX review longer than the implementation. It's been 18 hours. Nobody has used it.
I gave my AI agents a simple task: convert recipe URLs into a grocery list. Three hours later — database schema, Telegram keyboard, and a TTL cleanup cron.