Jan Siml is an AI architect at Swap Commerce and the creator of PromptingTools.jl, an open-source toolkit for integrating language models into Julia programming workflows. He builds AI systems around a demanding practical test: whether they help people complete specific tasks and produce measurable business results.
Siml studied at the University of St. Gallen, applying machine learning to investment-portfolio optimization in his thesis, and subsequently worked in data and product insight at LexisNexis UK before joining Swap Commerce. At LexisNexis, he developed generative-AI applications that translated internal data into recommended actions for sales teams.
He introduced PromptingTools.jl in 2023 and expanded it to support structured extraction, embeddings, classification, multiple model providers, and tool-based workflows. His companion project AIHelpMe.jl grounds coding assistance in Julia documentation, while his Julia community writing explores retrieval, developer assistants, evaluation, and model economics. He also created the Julia GenAI Jam podcast.
- Build for one valuable workflow. At AI Engineer World’s Fair 2025, Siml described how a small internal team built a sales-alert system that he said contributed several million dollars in annual recurring revenue. Teams with proprietary workflow knowledge, direct user access, and meaningful outcome data can build narrowly focused products that generalized vendors struggle to match.
- Track revenue, not vanity metrics. Technical measures such as F1 and nDCG provide useful safeguards, but Siml prioritizes revenue attribution: tracing AI-assisted actions through the sales funnel to actual commercial outcomes.
- Deliver proactive recommendations. Siml favors daily digests, timely sales signals, and suggested next actions over products that require users to initiate every interaction through chat. His public reflections on recommendation systems also question unnecessary architectural complexity.
- Make agents practical and testable. Siml evaluates models and frameworks against latency, operating cost, reliability, and developer control. His comparison of agent frameworks favors PydanticAI’s typed interfaces and retry handling, while layercode-gym supports simulated conversations, latency measurement, and regression testing for voice agents.
