Join us on Wednesday, October 21 from 9:00am–10:00am. Most ML stacks are three or four systems stitched together, and the seams are where projects die: features that drift between training and serving, models that cannot be traced to the data that produced them, and inference that cannot meet a latency SLA. This session runs a single model through its entire lifecycle without leaving the governance boundary of your data — from feature definition to production monitoring.
We’ll cover the Snowflake Feature Store for shared, versioned feature definitions and the online store that serves the same features at request time that you trained on — the single design choice that eliminates the most common source of silent production error. From there: experiment tracking for runs, parameters and metrics; distributed training and hyperparameter search on GPU compute pools; and open-source model integration via HuggingFace. We’ll walk through the Snowflake Model Registry for versioning and promotion, and the audit trail that ties a prediction back to the model, the code and the data that produced it. The session covers both inference shapes — high-throughput batch scoring and real-time model services for fraud, pricing and personalization — with the architecture and cost tradeoff for each. We close with ML Observability for drift monitoring and alerting, and realistic migration patterns from SageMaker, Databricks and homegrown stacks, including the Snowflake CoCo ML skills that automate the mechanical parts. Ideal for Data Scientists, ML Engineers, MLOps and Platform Engineers, and AI Architects.
Register now!
Topics Covered
- Snowflake Feature Store and Online Store
- Experiment Tracking and Distributed Training
- Snowflake Model Registry and Lineage
- Batch and Real-Time Inference
- ML Observability and Drift Monitoring
- HuggingFace and Open-Source Model Integration
- Migration from Legacy ML Platforms
- CoCo ML Skills
Speakers
Karthik Dulam
Senior AI/ML Architect, Applied Field Engineering, Snowflake
Register Here