I am primarily focused on the storage needs of RDS Open Source and Commercial databases. The most visible features I had led are RDS Optimized Writes, which delivers up to 100% improvement in performance for MySQL and MariaDB, and Dedicated Log Volumes (DLV), which reduces "long tail latency" IO events by 95% and improves IO performance by ~20%. I provided prototypes and "proof of concept", and technical leadership for the development and productization teams.

Contributed architectural guidance, technology ideas, rewrote badly performing libraries that lacked owners, other hands-on and leadership tasks Amazon's Principal Engineers are expected to do. My primary impact was on the server side of the product, but I had worked with all of the component teams to provide guidance on architectural coherence, algorithms, and design directions. I also focused on mentoring more junior engineers, prototyping novel features, and proposing forward looking technical strategy ideas.Invented a novel zero-click mobile and web app generation technology (see patent https://patents.justia.com/patent/11663199).

Contributed to client-side and neural net simulation software. Primarily focused on CS-1's high speed data I/O subsystem. Designed and implemented a high performance multi-tenant client-server session management and RDMA-like data transport engine using a custom binary protocol based on RFC 5041 and 5044 (C++ server, C++ and Python clients). Maintained and developed Cerebras' deep learning neural net simulation execution environment (C). Redesigned and rewrote (in C++) a flexible multiple datasets feeder service for machine learning neural net experimentation and profiling. Created a model of using Docker containers to emulate ASIC and FPGA components and embedded control service of Cerebras' deep learning engine. Built an internal Kubernetes registry and distribution service to deliver these containers to test systems. This allowed easy emulation of Cerebras CS1 system and accelerated development of drivers, verification infrastructure and tests ahead of custom silicon being available. Designed and co-developed a fast distributed data transport software to provide training data from client/feeder storage to the deep learning silicon (k18s, C).

Datera provided a scale-out intent-provisioned elastic iSCSI target. It was implemented as a cluster of commodity servers with tiered and all-flash attached, local storage. I had provided the architecture for Datera's distributed clustered storage system with some novel reliability features (improving on traditional RAID-1). I implemented large parts of the data distribution and shipping code in a Linux kernel driver (C), and created an extensive white box unit test framework (bash).

Provided architectural leadership for HDS' Hyper-Scaleout Platform (HSP): a scale-out, hyper-converged, shared-nothing Enterprise NAS and Big Data storage platform based on commodity off-the-shelf hardware components. HSP's storage stack inherited many high level requirements from ParaScale's PCS, but with a completely different architectural approach. After dissecting challenges of scaling metadata in a massively distributed system we built at Parascale, my design target for HSP was to drastically reduce the size of metadata. The team agreed on embracing the consistent hash approach to placement of data and services. This allowed me to design an extremely robust system with only cluster membership/topology being shared and distributed between cluster members. I designed the integrity recovery and storage rebalancing mechanisms to be "target-driven", which eliminated the need for central coordination of any of the cluster activities. A lot of this technology has been disclosed in patents owned by Hitachi. My primary "hands-on" focus was on the architecture, design and implementation of the distributed storage layer of the software stack (C++).

ParaScale's product, ParaScale Cloud Storage (PCS), was a highly scalable, self-healing, self-balancing, highly available, thin-provisioning storage software that runs on commodity Linux servers. It was trivially easy to configure, manage and maintain. It presented itself as an NFS, HTTP, FTP and/or WebDav target, and allowed easy "on the storage" application integration. The entire system ran in user space without any changes to the host kernel. I wrote large parts of the code for the distributed storage layer of the software stack (C++), designed the admin framework and wrote large parts of it (PERL), and implemented a large part of the test scripts. I also designed and implemented a simple error injection service for the test framework (a virtual filesystem "interposer" layer using FUSE and C).As the Chief Architect, I was responsible for guiding most aspects of Parascale's technical direction. Since this was a small company, I did what I expect everyone in a startup's technical team to do: whatever it takes to get the product developed and delivered. This included working with customers to get their requirements and feedback on the product, designing the high level architecture, detailed design of the product components for the team to implement, writing code, developing test suites, debugging, supporting QA, dealing with customer escalations, and so on.It takes a cohesive team to build a competent product, so being a team player and a leader, delegating work, mentoring more junior engineers, and picking up the slack where it exists are all part of the job.ParaScale's technology was acquired by Hitachi Data Systems.

VERITAS Software was acquired by Symantec in 2005. This position is the continuation of my VERITAS Software career.I managed an R&D team in the Storage Foundations division of the Data Center Management Business Unit (legacy VERITAS). This was an "advanced and new product technologies" organization and my team produced several new product initiatives and research projects, including novel storage layout research, file system scalability investigation, IO visualization, radically new file system defragmentation technique and tools, customer data anonymization tool for file system research, a novel disk failure prediction technology, and other areas of interest to legacy VERITAS storage products. Most of this new technology was protected by US and overseas patents, although some was released into public domain.My other areas of interest and involvement were:automated code quality tools (I was on Coverity's customer advisory board); continuous data protection (formerly on Revivio's technical advisory board, Revivo was acquired by Symantec);object storage technologies (co-chairing the snapshot sub-group of SNIA's OSD TWG);intellectual property and patent issues (member of Symantec's patent filter committee);open source (member of Symantec's open source and third party IP review board)

As a Distinguished Engineer (and a Sr. Technical Director), I took part in architecture committees and design/architecture reviews across all product lines at VERITAS. In addition to my primary responsibilities, I became a part of the intellectual property and patent program as a Patent Filter Committee and Open Source Review Board member; got involved in the source quality initiatives that culminated with licensing of Coverity tools; participated in SNIA OSD TWG; took an active role in the organization and content of VERITAS-internal engineering conferences; and participated in the customer meetings and conferences.

I was esponsible for architectural supervision and guidance of key projects of the APG division and got involved in cross-divisional architecture issues and external (OEM, reseller, customer) relationships.

Responsible for architectural supervision and guidance of VERITAS' flagship Volume Manager product (VxVM). In this capacity I focused on requirements and feature definitions papers to guide development of new VxVM features, and exploratory papers into various areas of interest for future VxVM directions. Also responsible for reviewing and guiding architectural development by groups within the VxVM organization, dealing with OEM and customer requirements and issue explorations, and participating in critical support escalations.

CVM started as a clustered version of VxVM commissioned by Sun. Over time it had evolved into a foundation technology for VERITAS Clustered file system, been ported to a number of platforms and integrated with several clustering frameworks. CVM evolved away from VxVM to become a separate, divergent product; later we integrated it back into VxVM, where it is now a license-controlled set of extensions. My role in the CVM group changed over time, as it grew from 3 people who didn't need a manager to over a dozen people in US and India, with me running the group.

VxFS: Worked on the early investigation of fileset cloning in VERITAS File System. VxVM port to SCO Unix: As a one-person "team", took the SCO Unix port of VERITAS Volume Manager from a questionable prototype to a product that had swept away the two dominant, well established rivals in SCO market place and won a "Product of the Year" award from a SCO UNIX publication.VxVM on SVr4: Worked on the 1.1 through 1.2 versions of VERITAS Volume Manager. The egalitarian structure of the group and the company culture had allowed me to be responsible for a number of conceptual and architectural inventions that had shaped the subsequent 20+ years of VxVM's evolution.

Locus was a pioneering single system image clustering company. I worked on storage management components and some kernel components of the AIX version of Locus clustered OS.

Office Automation software.

Wrote accounting and real estate-related applications.