Why Your ML Pipeline Needs an Agent (and How to Build One)
summary
Claims that an agentic ML pipeline reduces months of plumbing to minutes: a demo shows a five‑prompt chain (wine dataset) that produces ~150 lines of working notebook code, plots and a trained RandomForest in ~10 minutes versus ~7–8 hours normally. Uses Snowflake plus models from OpenAI/Anthropic/Gemini, runtimes like Ray, PyTorch, XGBoost, TensorFlow, and orchestration for distributed training, ML jobs and A/B serving. Cites 6,000 customers, millions of models, trillions of predictions per month and 10× developer productivity as reported outcomes.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 0:49Speaker intro
- 2:15ML practitioner time
- 4:41Context in agentic ML
- 6:25Unified ML agent platform
- 8:35Snowflake Koko internals
- 13:29Koko live demo
- 23:43ML agent architecture
- 25:10ML experiments to production
- 27:40Scaling personalized models
- 28:43Q&A
related talks
19mDatabricks
37mAutomated Machine Learning: Building a Pipeline
38mFantastic ML apps and how to build them
38mReal Time ML Pipelines in Multi-Tenant Environments
31mBuilding Agentic AI Platforms
01h 29mBuild Your Own Data Science AI Agents
27mScalable Data Pipelines - an ML Approach
10mML-based Log Analysis for Faster Debugging