Mardu Swanepoel

Mardu Swanepoel is Head of AI Engineering at Flinn.ai, developing AI systems that help medical-device manufacturers manage complaints, regulatory changes, and clinical documentation. His central concern is designing agents that understand professional workflows, explain their actions, and keep consequential decisions under human control.

Swanepoel studied mechatronic engineering at Stellenbosch University and pursued graduate research in electronic engineering. His connected bicycle-security prototype combined force sensors, signal processing, cloud connectivity, and smartphone controls to detect attempted theft.

He subsequently worked on Apple Music’s data-engineering team, helped develop decision-support software used by Apple and Move Inc., and joined Prewave’s data-science organization, building systems for supply-chain risk and sustainability monitoring. That progression through connected devices, data infrastructure, and risk analysis now informs his work on tightly regulated medical software.

At Flinn, Swanepoel builds around the documents, communication channels, and review processes professionals already use. His writing on user-centered medical-technology software argues that formal process diagrams often miss how decisions actually move among colleagues and tools.

Four principles for trustworthy agents

Swanepoel’s approach to effective agent design emphasizes:

• Focused agent modes: Restricting an agent to planning, research, or debugging narrows its action space, sharpens evaluation, and clarifies user expectations.
• Transparent execution: Visible task progress, consulted materials, and tool calls let users identify mistaken assumptions and redirect work early.
• Speed to understanding: Organizational context, remembered preferences, and professional playbooks help agents grasp how users actually want work performed before generating an answer.
• Reversible agent actions: Granular approvals, rollback, and familiar document-review tools limit the cost of mistakes and make higher-stakes delegation practical.

He also advocates structured coding-agent harnesses that reduce unnecessary intervention while maintaining production-quality software. For medical-device workflows, these safeguards shape practical applications including complaint handling, regulatory monitoring, and clinical-writing assistance.