Hi, I'm Alexander, building self-improving AI systems for business workflows

I'm Alexander Liss, Head of AI based in Denver. I've spent about 20 years helping companies through waves of technology, from social media to cloud, machine learning and now generative AI. These days I build systems that learn from their own outcomes. My stack is Hermes, Claude, Gemini, LangGraph, Claude Code, Codex and GitHub Copilot.

The most concrete result so far is a harness engineering field test: a harness that taught a cheap open-weight model real back-office work from its own mistakes. It gained +0.18 credit on unseen AutomationBench tasks at 15% of a stronger model's cost. It was preregistered and every number can be re-derived. I write about this on Substack, and I'm still working out how far the approach generalizes beyond that benchmark.

What I'd love from this community: feedback from anyone running agents in production on how you test changes to an agent's playbook before shipping them. I'd like to compare notes.