Why AI Is Never Really Done ft. Maria Vechtomova
Not every AI experiment should make it to production, and even when one does, deployment is far from the end of the story. In this episode of Bringing Data and AI to Life, host Nick Dobbins sits down with Maria Vechtomova, Co-Founder of Cauchy, MLOps expert and O’Reilly author, to explore what happens after AI leaves the experimentation phase. Maria breaks down the principles that have made MLOps more mature and repeatable, including traceability right through reproducibility as well as reliability and observability. What You’ll Learn Why some AI experiments should never reach production The four foundational principles behind strong MLOps practices How Generative AI and agentic systems change traditional deployment patterns Why evaluating AI output is harder when the definition of “good” keeps changing How the acceptable level of AI reliability depends on the cost of getting something wrong Why code-based evaluators, LLM-based judges, and human review all have a role in monitoring Why treating AI as “finished” after deployment creates long-term problems How enablement teams and better integration practices can help enterprises scale AI faster If you enjoyed this episode, make sure to subscribe, rate, and review it on Apple Podcasts and Spotify. Instructions on how to do this are here. This podcast is brought to you by Informatica, the leader in enterprise cloud data management. Links: Maria Vechtomova on LinkedIn: https://www.linkedin.com/in/maria-vechtomova/ Cauchy Website: https://cauchy.io/ Nick Dobbins LinkedIn: https://www.linkedin.com/in/nick-dobbins-guiding-data-modernization-with-responsible-ai/ Informatica Website: https://www.informatica.com/ Episode Highlights: 00:00 Welcome to Bringing Data and AI to Life 00:34 Meet Maria Vechtomova: MLOps Expert and O’Reilly Author 01:49 Every AI Experiment Shouldn’t Reach Production 03:17 The Core Principles Behind Reliable MLOps 05:20 What Changes: ML to Generative and Agentic AI 07:15 New Deployment Patterns for LLM Workflows and AI Agents 09:56 Why Trusted Data and AI Evaluation Go Hand in Hand 12:28 What Does “Reliable AI” Actually Mean? 14:12 How to Adapt AI Systems Without Rebuilding Everything 16:15 Why AI Still Needs Human Involvement 16:36 The Mistake of Treating AI Like Static Software 17:23 How Enterprises Can Scale Successful AI Experiments 20:05 Rapid-Fire Questions with Maria 21:20 Closing Thoughts