How to Build AI that actually Ships in Production - Aleksandr Kim

In this talk, Aleksandr Kim, Senior Data Scientist at Intuit, shares his expertise in building AI-powered features in production from fine-tuning BERT models in cyber security to engineering scalable data verification platforms. We explore the reality of moving beyond messy research code to build observable, cost-effective AI agents and automated pipelines.You’ll learn about:- Translating traditional machine learning metrics into actionable business outcomes- Validating large language model behavior through robust evaluation and alignment techniques- Pivoting from a generic chatbot project to high-value Slack automation workflows- Structuring outputs and guided reasoning layers to eliminate trivial AI summaries- Defining the overlapping skills between AI engineers, data scientists, and full-stack software engineers- Implementing multi-LLM routing logic and token caching to minimize enterprise API expenses- Identifying critical data infrastructure bottlenecks to determine when to pivot or drop an AI pilotTIMECODES:00:00 AI Engineering Production and Scalability06:12 Intuit Ecosystem and QuickBooks Products12:17 Aligning ML Metrics with Business Outcomes18:52 AI Engineers Conducting Customer Interviews25:13 Structured Output and Guided Reasoning31:13 Defining AI Engineering vs Software Engineering37:20 Cost Optimization and Multi LLM Routing43:26 UI Trends and Token Management in Industry49:33 Future Career Trends in AI Engineering55:46 Data Infrastructure Bottlenecks and ML FailuresThis session is designed for mid-to-senior level Data Scientists, Machine Learning Engineers, and Software Engineers who want to develop a highly practical, production-first approach to generative AI. It is especially useful for technology leads focused on reducing token overhead and building self-correcting agentic systems.Connect with Aleksandr- Website - https://alexkimds.github.io/- Linkedin - https://www.linkedin.com/in/aleksandrkim/