Across systems, AI products, and research, I have repeatedly identified a consequential technical direction, built the evidence behind it, and helped turn it into an adopted capability.
Establishing interaction tracking as a first-class perception target
I built Palona’s first lightweight interaction-tracking model, raising accuracy from under 50% with a flagship VLM to over 80% while reducing per-frame latency from 15 seconds to 0.22 seconds. The next challenge is turning abundant but largely unlabeled customer video into reliable, customer-specific interaction models—an ongoing research and engineering effort.
Read moreFrom a two-person experiment to an organization-wide coding-agent product.
I co-founded Rovo Dev CLI as a two-person project around the vision that the CLI would become a new interface for coding agents. When external open sourcing initially faced resistance, I helped build an internal open-source community instead. Within months, the project reached more than half of Atlassian engineers, with adoption comparable internally to Cursor, helping establish the case for an external-facing product.
Read moreSelf-supervised semantic discovery adopted across Google Ads, Search, and Cloud.
DeepClue combined an autoencoder with contrastive learning to discover semantic structure without predefined domain taxonomies. The same representation-learning approach was applied across Google Search, Google Cloud, and Google Ads—improving test coverage, retrieving more relevant context for LLM-assisted code generation, and reducing redundant integration workloads. Systems optimizations improved performance by approximately 50×, enabling the approach to scale beyond the initial research prototype.
Read moreChanging Google’s view of distributed-database performance
I benchmarked Spanner against Bigtable and showed that, for representative read-heavy cloud workloads, Spanner could deliver lower latency while providing stronger transactional guarantees. The work helped shift Spanner’s perception from a specialized database accepted despite its performance tradeoffs. Today, Spanner is the default database choice for a broad range of applications. I also simulated three proposed transaction features in a TPC-C-like benchmark, demonstrating the potential for approximately 3× higher throughput.
Read more