Tech Lead in Google Research working on applied graph ML and ML infrastructure:

• Data curation for labeling tasks

• Semi-supervised learning, label propagation, graph neural networks

• Large scale infrastructure for graph ML orchestration and MLOps

• Data infrastructure for Gemini pre-training

Applying R&D for a number of application areas including:

• LLM training and inference

• Recommendations and ranking

• Key member of the engineering team involved in building cutting-edge features, end-to-end product development, performance engineering and driving process change

• Built a generic protocol-agnostic socket engine that can multiplex client connections

• Re-architected the RPC pipeline to scale to 1000s of connections. This had a direct impact on revenue by targeting some of our biggest customers and helping them scale their datacenter deployments

• Major contributor for 'Synchronous Replication with Transparent Failover'

• Designed and implemented a distributed online integrity checker for synchronously replicated datastores.

• Designed and implemented a server side token framework to identify stale operations in the system and enforce server side access control

• Added support for synchronous replication of snapshots and clones involving arbitrary directory namespaces

• Paper titled "Logical Synchronous Replication in the Tintri VMstore File System" published in USENIX FAST 2018

• Worked on RAMCloud, an ultra low-latency DRAM-based distributed data center storage system

• Paper titled "The RAMCloud Storage System" published in ACM TOCS 2015

• Major contributor to designing and implementing scalable and low-latency secondary indexes

• Paper titled "SLIK: Scalable Low-Latency Indexes for a Key-Value Store" published in USENIX ATC 2016

• Master's thesis on managing objects and secondary indexes in RAMCloud

• Interned with the Tegra power and performance team.

• Implemented remote processor communication (RPC) using shared memory from scratch using upstream frameworks - rpmsg, virtio, remoteproc.

• Implemented a I2C proxy using rpmsg to relay I2C messages between the Linux I2C core and the I2C hardware blocks through the remote processor.

• Interned with the Advanced Technology Group (ATG)

• Worked on an efficient caching scheme for VDI boot-storms

• Leveraged block-sharing information from the storage filer, in the hypervisor

• Built a prototype on KVM/QEMU

• Achieved up to 90% bandwidth reduction and increased VM density

• Was granted US patent 8954718