I'm Cyril André, a Software Architect with 30 years of experience. I'm building Korvany, an AI-powered platform that transforms vague business ideas into development-ready backlogs. Along the way, I share everything I learn about building production software with AI. 🧵👇 – at Oudergem, België

The more I use AI in my software engineering workflow, the more I spend time to retain it from going too far. It becomes so obvious now that the intent, the actual need is something the AI doesn't understand or at least forgets very quickly and focused on technical or architectural "perfection". I feel like a dad trying to RE-focus his ADHD son.

I'm building Korvany, whose engine is AI-based. After a few iterations, I had to rethink what regression means and how to test for it. Regular regression tests might pass, but what if the model interprets the user's intent differently? We're not in a deterministic world anymore. I’m still figuring out what the right testing discipline looks like for that.

I've spent 30 years fixing failing tests. AI is teaching me a different habit. When an agent gets a red test, I don't immediately let it fix and retry. I stop and ask what the failure is telling me before the AI changes the evidence. – at Oudergem, België

This week I fixed a "bug" in Korvany. No code mistake. A user gave 2 statements. Natural for a human, but the engine stored them as 1 fact. Then one was confirmed and the other clarified... I was stuck. Had to split 1 answer into 2 facts. – at Brussels, Belgium

AI engineering note: When an AI suggested an implementation that seemed correct to me, I focused on understanding the reasoning behind. Now I'm looking for something I can verify independently. Another model approval is not an evidence; a test showing the result is as expected, is. Trust should not come from models but from guarantees we build around them. – at Brussels, Belgium

Architecture Note: This week I removed an AI call from Korvany. The UI already knew exactly what the user meant. So why ask an LLM to infer it again? Use AI where interpretation is required. Keep deterministic things deterministic. – at Etterbeek, België

I'm starting to think that AI won't reduce the importance of software engineering. It will simply move its center of gravity. Less effort spent writing code. More effort spent defining what correct software actually is. That's one of the biggest lessons I've learned while building Korvany.

7/8 Maybe the real assets are becoming: - the intent - the architecture - the invariants - the contracts - the tests that express those guarantees Everything that tells AI what must remain true.