Database Seminar - Jay Chia, Rohit Kulshreshtha
November 9, 2026 4:30PM—5:30PM
Location:
Virtual Presentation - ET
-
Remote Access - Zoom
Speaker:
JAY CHIA and ROHIT KULSHRSHTHA
Robotics produces high-volume time series alongside video and lidar, with different sampling rates, large payloads, and expensive GPU-based feature computation. These workloads shape the storage and featurization systems we are building at Eventual for pretraining-scale video/sensor data curation. We’ll show how standard open-source building blocks—Apache Iceberg, Parquet, object storage, and MP4—combine with careful database engineering to support these workloads at the scale of millions of hours of robot logs. We’ll cover data layout, temporal alignment, incremental feature maintenance, recovery from partial failures, and consistent reads across sensor data and model outputs.
—
Jay Chia is cofounder at Eventual, building data curation for physical AI. Prior to Eventual, he led teams that built ML/data infrastructure for multiomics biology at Freenome and for self-driving at Lyft. Eventual builds Daft, an open-source engine for multimodal data processing, which powers the overall Eventual platform for curating video and sensor data at the scale of millions of hours of robot logs.
Rohit Kulshreshtha builds the storage and metadata layer for robotics and sensor data at Eventual: transactional catalogs, change data capture, and moving hundreds of terabytes through featurization pipelines. Before Eventual, he was the architect of the Pinecone vector database from seed through its first 100 customers, designing its storage engine and blob-storage-backed write-ahead log. He also worked on the Hydro dataflow project in Joe Hellerstein's group at UC Berkeley, and on database systems at Snap and Amazon.
This talk is part of the Whatever Andy Found Interesting While Out At Sea
In Person and Zoom Participation. See announcement.
For More Information:
https://db.cs.cmu.edu/events/andy-at-sea-daft