5 Questions With Siva Gurumurthy

"5 questions with" is a chance for us to connect with leaders we admire and learn more about them and their work.

In this edition, we chat with Siva Gurumurthy, Chief Technology Officer at Harvey. Prior to this, Siva was the CTO at Motive and Director of Engineering at Twitter.

First, I have a deep conviction that the whole software ecosystem has changed. We’re no longer building to improve the efficiency of the user—we’re building to perform the work for the user. This is the software the majority of the new generation of engineers must understand how to build, and Harvey is a leading company in this space.

Second, very few AI companies have reached enterprise maturity. Harvey is the first of its kind native enterprise AI company that has achieved scale, and there are many enterprise-specific engineering challenges that come with that scale.

Third, the people. To have durable, long-term success, you need people with low ego and high growth mindset. Winston and Gabe (Harvey’s founders) exude that personality, and hire people who share those attributes. That’s the recipe needed for success.

Our goal is to ensure a substantial portion of repetitive and mundane legal work is performed by Harvey so that our customers can focus on the high-impact strategy and analysis that matters most. Therefore, our priorities are:

Team-wise, we’re thinking about global expansion. We operate in 50+ countries, each with its own flavor of legal regulations and go-to-market. We’re looking at the markets where we have a significant pool of customers and will eventually build teams that are focused on those markets. For example, we just announced expansion into Bengaluru, which we believe will be one of our top markets due to how rapidly the Indian legal market is evolving.

Obviously, hard skills and technical prowess matter. The ability to navigate complex requirements and break those down into concrete projects is essential, and that’s something we’re easily able to define through the interview process.

When it comes to soft skills, I prioritize two attributes.

First is a high level of ownership, not just over the code you write but over the business. I ran a team of over 200 engineers at Twitter and over 1,000 at Motive. The most successful people cared not just about their own professional advancement, but about helping the company win. We encourage this at Harvey by providing a lot of autonomy and internal mobility for engineers, and have created a flat and merit-based culture where the best idea wins, not necessarily the most senior or loudest person.

Second is low ego. You can’t be afraid to take a big risk and fail, or accept that you’re wrong, or to find a new solution that negates work you’ve previously done. Sometimes you’re working on something for months, and then conditions change—for example, OpenAI launches a new model—and then you have to throw away a lot of what you already did. To be successful, you have to be okay with that and stay focused on the customer problem that needs to be solved.

In terms of technical challenges, here’s some of what we’re thinking about:

One thing that’s technically unique about Harvey is that we are one of the world’s largest document processing systems. A single enterprise customer might upload anywhere between 200,000 to over 1 million documents and then ask our AI agents to perform actions on those documents.

That’s really like performing 200,000 to 1 million actions. That creates a fan-out problem where one action becomes hundreds of thousands of actions. This is a very difficult distributed systems scaling problem in the AI agent space.

Harvey’s culture is really customer-first. Our engineers partner closely with our customers and are empowered to translate their problems into technical projects, then get those solutions into production. This is an exciting environment to work in for engineers who want to get the maximum reward for customers, quickly.

When we first present Harvey to customers, we always start with the lowest-common denominator set of use cases. From there, customers ask us if we can actually do X, Y, or Z new types of tasks, and that informs how we evolve the core platform.

One example is a product called Playbooks where customers can upload their organization's contract standards—let’s say, restrictions on what language is accepted that can be applied to future contracts to determine which fit those restrictions.

Some customers said, “Well, we don’t know have a playbook.” So we had our engineers come in and figure out how to reverse engineer a playbook based on the customer’s existing contracts and show them, “Here’s the playbook you’ve actually been using this whole time.”

We took that use case and moved it into the core Harvey platform even though it started as a custom project for one customer. That’s an example of how we work to define the product on the go, to find product-market fit, and how engineers can have outsize impact at Harvey.

The most common by far is, “Why is your world different from foundational model companies? Why couldn’t I just go to ChatGPT, ask the same question, and get the same result?” That’s an important question to address, and the answer is multipronged.

First, think of any operating system—Apple’s iOS or Google’s Android, for example. There are apps built on top of those systems that are $100B companies. Technically speaking, Apple and Google could have built any of those apps because they own the foundation. What made those app companies successful is focusing on a specific domain and problem, then building what is needed for those customers. The same logic applies to Harvey and ChatGPT. Verticalization is necessary for the best results.

Second, the enterprise needs more than just a raw foundation model. You need corporate governance, security, privacy. You need to make sure the software plays well with other software and integrations. These are the considerations at the core of how you build true enterprise software.

Third, distribution. We care a lot about owning the mental model for legal AI. We want Harvey to be the name people think of even if 30 other competitors emerge in this space.

Fourth, data. We work on a specific vertical and understand it deeply. As a result, we’re able to collect data that can help fine-tune our foundational model and build a post-training framework on top of it that creates more accuracy in future legal work.

Fifth, the technological advantages. We have higher per-product requirements than a foundational model would. A bit of hallucination is okay with a general consumer product. But in the legal space, you have to be 100% accurate. You have to have the right sources and the right citations. There is extremely low tolerance for hallucination, which is a unique technical aspect of how we build.

Those are the major advantages I point to.