Nov 2017 – Dec 2018 · 1 yr 1 mo · Redwood City

Nov 2016 – Nov 2017 · 1 yr · Menlo Park

• I worked on the data infrastructure of the Lucid Air. Everything from collection, storage, analysis of data, learning, deployment of infrastructure, and data visualization. This includes raw data from the car's sensors as well as autonomous driving data.

Aug 2016 – Nov 2016 · 3 mos · Palo Alto

• This is the nucleus of the tech teams, we are both builders and breakers - developing the technology and systems to maximize developer efficiency and deconstruct limitations on scalability, performance, and reliability. This includes (but not limited to) developing scalable benchmarking solutions as well as the proper tooling to support the growth of the platform
• I worked on datacenter-wide optimizations at the kernel level as well as PHP application level improvements.

Jun 2015 – Sep 2015 · 3 mos · Menlo Park

• My work at Facebook focused on scaling the largest web capacity in the world, by overhauling the benchmarking capabilities of the entire infrastructure. I designed, implemented, a more flexible way to collect and combine system and application level metrics. This is currently running on the entire Facebook fleet. Of course, with incredibly low CPU utilization. Finally, through building this tool I collaborated with other teams to remove unnecessary dependencies from the performance sampling software.

Jun 2014 – Sep 2014 · 3 mos · Santa Clara, CA

• My work at Intel focused on research direction that would help improve datacenter efficiency. Nowadays, Solutions range from better air conditioning strategies, to using virtual machines that can migrate tasks to certain servers in order to power down others, to fine grained power management systems that enhance energy proportionality. The project we proposed was in the application scheduling and management arena, which is one of the largest topics of interest within data centers nowadays. In my intership, I looked at the solutions, problems, and tradeoffs that competitors, academia, and Intel, have experienced in the realm of scheduling and management of application workloads. Specifically I looked at determine optimizations to make at the processor level to improve performance of datacenter applications through scheduling logic, specifically tailoring those that would support long-term business strategy.
• At intel I lead the genesis of the datacenter project, configuring and deploying scripts to manage 1000 cores on 12 machines. I then developed a platform in Ruby that allowed custom made tests to run and benchmark the infrastructure. This allowed for datacenter-wide testing suites that would collect, parse, and present experiment data.
• Final I nalyzed performance metrics of Hadoop, Memcached, and other datacenter workloads. Focusing on memcached performance at scale and the numa performance impact in 95% QoS.

Apr 2013 – Jun 2016 · 3 yrs 2 mos · Stanford

• At Stanford I worked on several research projects:
• Memcached Optimization: Paper published at Workshop on Resource-Efficient Cloud Computing @ ISCA
• Here is the abstract of the paper:
• Latency-sensitive applications are widespread in datacenters and support services such as social networking, search, and commerce. Methods to protect quality of service have created inefficiencies in datacenters that could be ameliorated with a latency-violation predictive mechanism. We present such a system, Parachute, that achieves good performance prediction across a variety of load characteristics, including queries per second, differing write-to-read ratios and request key and value size distributions.
• Another project I worked on was an ARM 64 chip benchmarking on APMs Mustang board:
• The goal of this project was to use a prototype server with an ARM64 chip and optimize datacenter workloads to use novel features in an Arm64 board. The goal was to show significant latency, throughput or power efficiency improvements. As part of the project, we also benchmarked a lot of existing apps on these new chips.
• Motivation
• The large-scale data centers used by companies like Facebook and Google are power limited. Hence, there is a lot of interest to use power efficient technologies for processors, memories, networks, and storage.
• We wanted to test the performance of the Mustang board to determine its relative performance to other servers in the industry. Performance is measured in three basic pillars, CPU, memory, and network, and the analysis is scaled to the power consumption of the server.

Apr 2013 – Jun 2016 · 3 yrs 2 mos · Stanford

• This research focused on developing a computer system capable of automatically recovering small aircraft (fixed wing drones or UAVs) from the dangerous and statistically significant Stall/Spin upset which, if implemented correctly, would have a notable impact on the safety of small personal aircraft and the training requirements for pilots.
• This research project is in essence an interdisciplinary endeavor, since it required deep understanding of 3 distinct fields: Aircraft Aerodynamics, Control Theory and Artificial Intelligence. Not only does it require deep understanding, but also the ability to precisely identify and combine those concepts that will make possible a breakthrough both at the theoretical and practical level of Aerospace Systems.
• I worked with a PhD student on the implementation side of the research topic. Developing this system required doing extensive flight testing on airplanes ranging from 3-12 foot wingspan. I operated all the radio controlled aircraft needed to carry out the research, including putting and recovering them from different stall/spins and other maneuvers. Furthermore, I also provided building and setup support when assembling these complex systems. Finally, I collaborated in developing the machine learning algorithms and the feature selection process of the stall spin detection.

May 2011 – Apr 2012 · 11 mos