
The Gradient: Perspectives on AI Shreya Shankar: Machine Learning in the Real World
6 snips
Sep 7, 2023 In this episode, Shreya Shankar, a computer scientist pursuing her PhD in databases at UC Berkeley, discusses the challenges faced by machine learning engineers, including ensuring consistent performance and accuracy in production ML models. She also explores the importance of preserving information for anomaly detection, quantitative statistics for data cleaning, and the evolution from MLOps to LLMOps and FMOps. The speakers discuss the difficulties of implementing ML models and ML monitoring, as well as varying perspectives on AI as a product and its impact on customer expectations.
AI Snips
Chapters
Transcript
Episode notes
First ML Engineer Spent Months On Data Plumbing
- Shreya left her ML engineer role after spending a year building data integrators and cleaning pipelines instead of training models.
- At Viaduct she was the first ML engineer and found data collection and cleaning consumed months before any modeling could begin.
Treat Models As Materialized Views Managed Like Data
- Shreya maps database principles to ML and highlights separating logical from physical concerns as crucial for scalable ML systems.
- She argues models are frequently materialized views and data-management ideas like incremental view maintenance should inform MLOps design.
Make Observability A Three Step Roadmap
- Build observability as a three-step roadmap: detection, diagnosis, and reaction.
- Log sufficient runtime information, maintain searchable logs and summaries, and prefer simple guardrails or rejection rules before retraining models.

