Dan Feng describes Maven Clinic’s AI transition as three connected changes: internal tool adoption, AI features in products and changes to how people workWorkflow automation uses software or AI to complete a connected sequence of routine steps with less manual effort.. His approach gives early adopters room to experiment while making shared infrastructure easier for the wider organization to use. Supporting different tools and listening to concerns are part of the adoption process.
Dan Feng says AI implementation tools change the skills his team values, increasing the importance of independent problem solving, product understanding and architectural judgment. He favors a broad long-term direction alongside concrete two-to-four-week delivery cycles and short requirements documents that can change as the team learns. These are his organizational practices, not evidence that every project benefits from shorter planning.
Dan Feng explains that greater code output creates a review bottleneck. Maven Clinic began with lower-risk, easily checked tasks, then expanded AI assistance while retaining engineers’ responsibility for architecture and evaluation. His account includes smaller pull requests, stacked changes and avoiding approvals that merely create false confidence. Fully automated development and production repair remain goals rather than completed capabilities.
Dan Feng distinguishes failures by their consequences, contrasting a retriable appointment-scheduling error with an incorrect reimbursement amount. He describes cross-checking receipt extraction with different models and escalating uncertainty to a humanHuman judgment in AI is the accountable interpretation and decision-making people contribute when setting goals, evaluating evidence, and managing consequences.. Repeated integration testsAI regression testing reruns preserved cases to detect whether a model or agent update has broken behavior that previously worked., automated conversation evaluationEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria. and manual review provide additional feedback. These are described operational safeguards; the talk does not establish clinical efficacy or prove that agreement between models guarantees correctness.
Watch on YouTube




