We believe focus is of critical importance for a startup studio, even more so than for a venture capital firm. A startup studio’s worth, first and foremost, is defined by the value its entrepreneurs place on the studio’s ideas and validation processes. In our experience success is vastly improved by building deep competencies in few vs. many technology domains.
With this end in mind, early last year we researched the landscape of compelling technical, consumer, and enterprise trends and drew inspiration from Madrona Venture Group’s 2017 and 2018 investment themes. From there, we considered our team’s strengths and experiences and the inherent advantages of the Seattle tech ecosystem. Our final step was socializing and refining our thinking with our investor network, 20 of the top seed and VC firms in the country. This post covers our core focus area ML/AI, and future posts will explore frontier tech investments we are making.
Vertical Machine Learning (ML)
In the venture-backed startup world, machine learning (ML) is ubiquitous. Nearly every startup that walks through venture capitalist’s doors has an ML story. In all technology sectors, trends ebb and flow and ML is no exception. These days, vertical ML is the place to be. So what are the key components of a sound vertical ML startup? First, it might be worthwhile to talk about what makes a Horizontal ML company:
The term horizontal in the context of products or companies typically denotes applicability of technology to multiple industries and problem spaces. Given the common data and analytical patterns required for solving supervised learning problems, many ML platforms can be fairly generic and applied across many industries such as Healthcare, eCommerce, Finance, Travel, Agriculture, Marketing, etc.
But the intelligence of these horizontal ML platforms has yet to reach a level of automation that spans domains, where they can simply be pointed at an arbitrary data source and be trusted to do the right thing. It turns out that the devil is often in the details when it comes to developing analytic solutions for specific problems. Domain experts are still needed to understand the business problem, acquire/analyze/engineer the data, and train the models to assure the problem is being solved in a rigorous fashion.
Conversely, Vertical ML is about data, technology and expertise applied to solving a problem in a specific business vertical or domain. Domain-specific data, experts and customer-facing product are all aligned towards solving a single problem. There can be proprietary ML technology too, but off-the-shelf algorithms and techniques often suffice. It is the application of the analytics technologies applied to a domain-specific problem that is novel. We should remember too that cutting-edge algorithms and machine learning technology tends to become commoditized quickly. The competitive advantage shifts to understanding and owning data and engineering it into a form that maximizes the extraction and identification of high business value signals . The delivery of these signals or predictions in the product UX can also be highly customized and tailored to the user’s background and knowledge.
Conversely, Vertical ML is about data, technology and expertise applied to solving a problem in a specific business vertical or domain.
A great example of a vertical ML company was Farecast (sold to Microsoft), a Madrona-led startup where we led product and technology. Founded by Oren Etzioni based upon his research at University of Washington and originally incubated inside Madrona, Farecast was an innovative analytical consumer-facing product for the travel vertical. Farecast’s core differentiated value was predicting the direction and magnitude of future airfare price fluctuations (set by airline revenue management systems), giving the consumer a new level of transparency on ‘when to buy’ their tickets. Although we leveraged state-of-the-art machine learning technology for computing the predictive signals that we delivered to our users, much of the challenge we faced was procuring and understanding the airfare data which powered our analytics technologies. Given the mechanics of airfare pricing data is not well known and actually quite arcane, much of our learning came from analyzing the troves of data we harvested every day. We eventually collaborated with travel domain experts who augmented our knowledge of the mechanics of airfare pricing engines.
Our user experience was also carefully crafted and honed over time to be appealing and familiar to consumer leisure travelers, while at the same time conveying these novel predictive signals we were computing with our tech. We could have just shown generic analytic visualizations familiar to many data scientists, but these would have likely alienated our non-tech-savvy consumer customers. Our focus on this specific analytics problem in the travel domain allowed us to create an approachable product for consumers that was also very accurate.
We know from our own experiences what makes a vertical AI startup, but we wondered about other perspectives on vertical ML. One of the best articulations we found was this article by Bradford Cross. In this great post and video he lays out 4 key characteristics of a vertical ML company:
- The founding DNA of the company combines subject matter expertise (i.e a domain expert) with technical expertise.
- The subject matter expert has experience in the vertical, has identified a priority problem, and has a rolodex of industry contacts to leverage for market and customer development. Technical expertise builds and tunes a specific solution to the identified problem
- Company offers a full-stack, end-to-end product (not just analytics pieces) — Owns customer UX, business/customer relationships
- Company procures/creates/derives strategically valuable proprietary data
- Has Machine learning technology that delivers core value to the business
The team at Madrona Venture Labs has a wealth of experience building Vertical ML startups. It is a key strength for our team and of deep interest to our investor network. After Farecast, Mike and Jay both went on to create more companies that leverage data and machine learning across domains such as eCommerce (Decide sold to ebay) and healthcare (Medify sold to Alliance Health Networks). Our design partner Jason Flateboe has extensive experience building domain-specific user experiences that are tailored to specific applications and customer types. And of course we believe many industries are still waiting to be transformed by leveraging data and machine learning.
Finally, Seattle is a great town to do machine learning startups. Between the University of Washington, Allen Institute for Artificial Intelligence, our native big tech companies (AMZN, MSFT) and the other big tech giants (Apple, Google, Facebook, eBay, etc) setting up shop with AI/ML focus and growing here, (and here!) plus all of the local startups. Seattle has a strong local network of machine learning talent! And of course Madrona has a strong history with vertical (e.g. Farecast, Decide.com, Placed, etc.) and horizontal ML/AI companies (e.g. Turi, Lattice, Algorithmia, XNOR, etc.), just to name a few.
What we are doing to continue to build our competency? To add unique value to founders, we are building up our capabilities through our team, ideation, and the Seattle community. An example, last summer we hosted our first Machine Learning Ideation Workshop in which engineers, designers, and product managers came together over a weekend to build six ML/AI company concepts. The winning idea is now being developed at MVL. Almost a year later, we just finished our second workshop. Once again we had great people, great ideas and a lot of passion for machine learning. We’re excited to continue working with the teams from this latest workshop to help them to continue to develop their ideas and align our passions for creating great ML companies in Seattle!