Benjamin Verbeek is an AI engineer at Lovable building continuous improvement loops that help coding agents recognize failures, repair their tools, and guide people without programming experience toward finished applications. Before moving into AI, he worked across particle physics, satellite design, and fusion research.
Verbeek studied engineering physics at Uppsala University, where his projects included statistical software for particle-physics experiments, spectroscopy of muonic atoms at Switzerland’s Paul Scherrer Institute, solar-cell research, and FLAMES, a proposed satellite constellation for monitoring wildfires. He also founded a physics and astronomy camp for secondary-school students and received the Anders Wall Foundation’s Uppsala Student of the Year 2022 award. At Novatron Fusion Group, he developed Langmuir probe diagnostics for studying plasma behavior.
He joined Lovable initially as its first data analyst before becoming a developer working on its core agent. His responsibilities have included frontier-model evaluation, product development, language-model spending, and improving the agent’s operating instructions.
Building agents that learn from friction
- Project completion as the real product metric. Verbeek evaluates coding agents by whether users actually finish and deploy applications. A programmer can often recover from a missing credential or broken configuration; someone without that background may abandon the project entirely.
- A continuously evaluated knowledge base. LLM judges identify stalled conversations, successful recoveries, and recurring obstacles. Related solutions are clustered into an internal Lovable Stack Overflow, selectively supplied to the agent, and compared against holdout groups. Guidance that stops helping is retired as models and product capabilities evolve. His AI Engineer Europe talk details this architecture.
- An agent vent tool. Verbeek built a mechanism allowing Lovable’s agent to report broken tools, confusing documentation, and platform failures directly to engineers. One recurring report exposed filename-copying bugs caused first by ordinary spaces, then by non-breaking spaces and other unexpected characters. The agent-feedback workflow now includes automated deduplication, investigation, and pull-request preparation, with engineers reviewing proposed changes.
