Research Insights Made Simple #31: AI4SDLC — what I would do differently

Where do you start building an AI stack: choosing a model, buying GPUs, or writing your own harness? In a lightning talk at Deep Tech Night on September 5, I suggested starting with the ownership boundary: what to rent, what to adapt to your environment, and what to keep under your control.

At 13:00 MSK (UTC+3) on September 22, I’ll present the director’s cut of that talk in a live stream. It received more than 100 votes in an earlier post.

The episode focuses on a few practical questions:

• Why write your own agent loop when the market already develops ready-made ones? And when is a custom harness justified?
• Why does the same model produce different results with different tools, context, and permissions?
• How do you distinguish a clean trace from a completed task, and verify that a change actually helped?
• Which tasks can move to a smaller model, and where should a larger one remain?

The talk slides are already on the website. The episode will be on YouTube on September 22. Join me, especially if you are deciding which parts of your AI stack are worth owning.

#AI4SDLC #AI #Agents #Architecture #PlatformEngineering #Evals