Roman Petrov

The main rule: anything that goes to the cloud stops being private. My approach is: agents are great, but only if they run locally (through Ollama or LM Studio). Yes, the model will be dumber than Claude or GPT-4, but at least I know my financial reports won’t leak into a training dataset. Trusting an API with bank account access without a "human in the loop" and strict limits is basically...

I started with a really basic thing - an internal report generator. Chat GPT plus a bit of manual coding. The hardest part was trust: I’m used to understanding what’s happening under the hood. Here, you end up reading and checking more than writing. Didn’t make any money, but it changed how I approach prototyping - now it’s actually really cheap

AI is great at automating tasks, but not responsibility. What’s most valuable today is being able to frame the problem, make decisions under uncertainty, and own the outcome. Code, design, and copy are just tools - but systems thinking, prioritization, and real business understanding still don’t automate that well. That’s what people will pay for