Rob Earhart Email & Phone Number

Seattle, Wa, Us

Burlingame, Ca, Us

My goal here was to work through my backlog of interesting project ideas, learn new fields and skills, volunteer a little, and generally enjoy life. :-)At Vertex and Intel, I'd gained a lot of experience optimizing ML systems, but never really had the time to grok the higher levels of the ML stack. So I've spent some time rectifying that, learning enough PyTorch to do what I want, training models, and experimenting with ideas and tradeoffs -- learning how to build software where test success isn't nearly as clear-cut as it is with systems programming, and trying to build up an intuitive understanding of what might work and not work in ML model design. (And really wishing I had a more rigorous understanding here, but I suspect that's an area where we as an industry have a lot of work to do.)

Santa Clara, California, Us

Deep Learning Software EngineerDeep Learning Software Engineering ManagerAt Intel, I incorporated the Vertex codebase into nGraph, and then worked on adding support for various other internal ML processors. I got to present demos at NeurIPS and MLSys, and helped port the system from our custom internal dialects to MLIR, which I still think is maybe the most elegant compiler framework I've had a chance to work with.As a manager, I led a small team validating that our OpenVINO compiler stack worked well with Meteor Lake's NPU, and then grew the team to integrate the compiler component with WinML. This was fun because WinML really wanted to be in control -- and from my experience with NT kernel, I definitely appreciated why! -- but the Meteor Lake NPU is a delightfully intricate little machine, and it's a little tricky to realize its full performance through an abstract API, so there was a lot of cross-company cross-team cross-versioning-boundary design work that needed to take place, while simultaneously growing the team and growing the people on the team.

Seattle, Washington, Us

Principal Software Engineer / Jack-of-all-trades (first non-founder at a startup => doing a bit of Everything)At Vertex, we built a polyhedral tensor compiler, and integrated it with Keras. This let us describe ML operations mathematically, and compile them to pretty decent kernels for arbitrary GPUs, with kernel fusion, buffer management, and kernel scheduling handled automatically. It was a pretty good system for quickly evaluating accelerator tradeoffs, which led to our being acquired by Intel.

Mountain View, Ca, Us

Software engineer, Google Compute EngineTechnical lead & co-founder, Google virtual machine teamArea lead, Cloud Billing systemTechnical lead, Google ExacycleSoftware engineer, Google DataflowSoftware engineer, Google APIs teamMy 20% project grew into the virtual machine monitor running Google Compute Engine--pretty cool, and my introduction to being a tech lead for a project, which was tremendously fun. This job was also my introduction to distributed systems at scale; also a very fun area, although also an area where I wish we had better tooling; I'd love to make building a distributed system as easy as laying out relationships in a spreadsheet.

Redmond, Washington, Us

Software engineer, NT kernel & low-level NT usermode libraries.Software engineer, Microsoft Hyper-V; owned the hypercall, guest physical address translation, and memory manager layers, as well as the WinHV driver.Software engineer, Midori; microkernel and low-level usermode development.I love writing kernel code--it's actually surprisingly easy, but it requires thinking very carefully through interesting intricate details (which states the processor might be in, which cache lines you're writing to, &c), and requires getting your abstractions really Right, so that you can simplify the code as much as possible and make it obviously correct. It's definitely easy to spend too much time architecting and abstracting, but the kernel teams at Microsoft were a great place to learn the tradeoffs and see the outcomes in both directions.

Pittsburgh, Pa, Us

Worked for Carnegie Mellon Computing Services; general campus computing support.Software engineer, ACAP/IMAP protocols & serversSoftware engineer, user account systemSoftware engineer, user help information systemSysadmin, UNIX & AFSThis was my "working my way through school" job -- staff got free classes, although it took me twice as long to graduate as I'd have liked. It was a fantastic place to be: we were close to the forefront of scaling computing services to a campus population, tackling a wide variety of problems, and trying to imagine what the world might look like in a distant future when everyone might have a portable wirelessly-connected computer.