At Bubble 42, I evaluate agentic coding trajectories to understand how LLM agents use tools and where they fail. I use the Harbor framework and opencode agent in Docker, and document how prompt variations affect output quality.
At Unilever, B&W and PC Factory, I built an IT Support Hub that routes tickets with a RoBERTa classifier achieving 80%+ accuracy. At Unibotech, I built real-time pipelines from PLC signals into dashboards and AI model inputs, and developed a RAG assistant for industrial equipment documentation.
