TL;DR: Why I am joining H2O. This is long and rambling, and mostly written for myself so I can review it in a year or two.

Today I start a new role at H2O, a machine learning applications company with a open source platform for data analysis. Three years ago I wrote a blogpost about why I chose to work at Sift Science. This blogpost is similar: this is my hypothesis for why H2O is an interesting opportunity. I thought about it from the following lens:

• Company Mission
• Team
• Learning Opportunity
• Unique Contribution

I will focus mostly on my rationale, but I will also add a few pre-mortem notes.

Company Mission

As I understand it, H2O is a “tools” company. They focus on making really great machine learning software, which data scientists and engineers can use in applications they are building. The way I explained the difference between H2O and Sift Science was this: if Sift Science created cars (i.e. a full featured product designed to solve a specific problem for customers.) then H2O makes engines.

There are many companies with this business model – make tools available for free, charge for services and consulting. What makes H2O interesting? To me it is the company’s thesis about the future of data science and data analysis.

Today, in order to do large scale machine learning and data analysis, you must first understand statistics and modeling. On top of that, you must also understand distributed systems engineering. Why? When a data set is sufficiently large, it is infeasible to perform all the computation on a single computer. It’s simply too slow. The solution is to split up the task and have multiple machine perform the task at the same time i.e. distributed systems. This is why good machine learning engineers are so hard to find. A suitable candidate needs to understand both distributed computing and machine learning to do their job well.

What if that’s not necessary? Machine learning and data analysis does not intrinsically require understanding distributed systems. If we can remove the distributed systems engineering requirement, then large scale data analysis will become accessible to a lot more people. There are an order of magnitude more data scientists and statisticians out there than there are machine learning engineers. These are the people who will be enabled by H2O’s software to run analysis they had never been able to before.

Frankly, I can’t wait to see what creativity this will unleash.

Pre-mortem: I can see a few ways things might go wrong,

• Abstracting away distributed systems proves impossible, or the abstractions will prove too leaky. In this case, H2O’s software will still be valuable, but will not be paradigm-changing. The status quo of big data analysis requiring engineering+statistics “unicorns” remain.
• Abstracting away distributed systems proves too easy, and competitors are able to quickly replicate this ability. In this case, H2O will need a different competitive advantage.
• The value of data analysis pipelines resides not in better algorithms, but in better feature extraction, and therefore competing to be the best ML package is futile. Frankly, this is the one that actually worries me most. ML packages can compete on speed and accuracy, but both of those traits have diminishing returns. As I witnessed at Sift Science, real world prediction gains comes just as often from improvement in feature extraction as from algorithm improvements.

Team

My first connection with H2O came through Erin Ledell. This was after R2D3 came out, so naturally talked a bunch about how to communicate machine learning insights to a broader audience. Through our conversation Erin’s passion for good open source software clearly shone through. She really wanted good tools to be accessible by as many people as possible. As I met more folks from the engineering team, it’s evident that this passion is shared by many at the company. The theme underlying every interview question was, “How can we make this software better? How can you help us make this better?”

In addition to a passionate and sharp R&D org, I was also intrigued by my conversations with Kanishk. He is the VP of Customer Applications, and we had an interesting and wide-ranging chat about how H2O’s software might get used. The most interesting insight from Kanishk was how algorithm speed opens up new frontiers. I was reminded of the maxim about how the abundance of a resource leads to more imaginative uses of it, and my mind danced with ideas about using machine learning for art.

It was clear that H2O has put together a really impressive team.

Pre-mortem:

• I may get silo-ed as a data-visualization designer in a engineer centric organization. While I am glad that H2O reached out specifically because I demonstrated that I can communicate machine learning concepts with R2D3, that doesn’t mean the whole organization understands the value of product design. I guess I was worried about this when I joined Sift too, but that worked out fine.
• I will have difficulty nurturing design culture there, since the culture of H2O is largely already set. I joined Sift Science when it was eleven people. It was much easier to have a personal relationship with everyone who was there, simply because I was there while the organization grew. I made it a point to explain the role of design to everyone when they started. My influence will be much more diluted at H2O, which already has 50 members.
• Everyone at H2O seemed super focused. One of the things I really enjoyed about Sift Science’s culture was its emphasis on work life balance. I didn’t really probe the culture at H2O as much as I should have. Other than a really smart and focused team, I am not sure what I am getting into culture-wise.

Learning Opportunity

I have an draft of a blogpost called “What comes after the Triple Threat Designer” where I contemplated what kind of skills I should develop next. My personal career thesis was to be fluent in design, technology, plus product strategy, and to create a unique blend of capabilities from those disciplines. At Sift Science I was able to prove out that this blend was in fact effective, especially in these two situations:

• Making design decision while taking all three perspectives into account.
• Vetting and refining ideas from various parts of the company into feasible projects.
• Translating ideas across functional boundaries effectively.

At H2O, I want to add data science to this mix. Thanks to working at Sift, I have developed a particular affinity for data products. I learned a lot about the in’s and out’s of machine learning, and saw how machine learning systems gets deployed in real world applications. At the end of the day though, I was designing for fraud managers and analysts. In reviewing my various hackathon projects at Sift, I found that half of my projects were tools I built for Sift’s ML team. These were often the hackathon projects I enjoyed the most. At H2O, designing tools for data scientists will be one of my primary roles. Given the range of customers that H2O has, I have the opportunity to see a broad range of data science teams and processes. My goal is to use this opportunity to develop a unique perspective for designing data products, and become a world class data product designer.

Pre-mortem:

• If I mess up this opportunity, it will because I get too caught up into building “cool” things, and do not spend enough time actually talking to customer and clients. I know I have the tendency to obsessively polish a project, which worked well for R2D3, but will serve me poorly if it takes away from learning from actual users.

Unique Contribution

There a lot of product designers out there. Many of them have front-end engineering chops to bring their designs to life. The pool starts to narrow if you want the product designer to have a data-visualization focus. Of these, very few are also familiar with machine learning and data science. H2O is looking for exactly this, and I am fortunate that my skill set and interests are such a great fit. H2O aims to be the go-to platform for open source machine learning. It has already developed a great set of tools algorithms, and a track record for solving real business problems with them. To broaden adoption, it needs to improve the usability of this tools, and help people understand their results through data visualizations. These are things I have done before, so I am confident that I have the right set skills to push H2O’s product to the next level.

Pre-mortem:

• This data+design+product synergy may be overstated if the needs of the company turns out to be more simple. e.g. if lots of individual data visualizations needs to be produced independently of each other, and customers don’t find it important that they are cohesive.
• This synergy is not scalable. This I already know is true. As companies get bigger, roles ought to become more specialized. As H2O brings in more specialists, folks with hybrid skill sets either move into management, or find their opportunities diminish. A company at H2O’s size should be starting to hire more specialists. I am not sure how this will play out.

The Point of a Pre-Mortem

Folks often complain to me about the morbidity of ‘pre-mortems’ – but to me that’s exactly the point. I am really excited about H2O, but I want to make sure I remind myself that failure is always a possibility. I don’t think I am a pessimist. I know most of these ‘pre-mortems’ likely will not come true, but it is important to articulate it before hand, so it can be a point of reference for review later.

As I was writing this post, I went and reviewed my pre-mortems for joining Sift Science. Essentially all of them were unfounded worries. I had a very happy two and a half years at Sift. The only one that was remotely accurate was the one about “wanting more variety.” It took 28 months before I even starting thinking, “oh maybe there are other interesting problems out there. I should go out and look.”

Anyway, here’s to a new beginning. Let’s see how this plays out.