The Center for Data Innovation spoke with Mike Flaxman, vice president of product development at Heavy.AI, a San Francisco-based analytics company that uses AI to help enterprises explore their geospatial data, turn it into actionable insights, and generate visualizations. Dr. Flaxman spoke about how geospatial data helps enterprises make risk assessments, Heavy.AI’s vision for AI-driven conversational analytics, and the challenge of prioritizing where to invest resources in his role.
Martin Makaryan: What does Heavy.AI offer?
Mike Flaxman: Heavy.AI started with a mission to expand access to big data. We offer a single software to query, analyze, and visualize geographic and spatial data, as well as derive insights that can help inform high-stakes decisions. We have partnered with Nvidia to use graphic processing units (GPUs), which allow us to perform queries 10 to 100 times faster than traditional central processing unit (CPU)-based databases. This high-performance capability is particularly useful for customers dealing with large, complex, and fast-moving datasets, such as weather models or telecommunications data. In addition to our querying software, we offer digital twins and scenario planning. These tools allow users to project future outcomes and evaluate potential courses of action based on geospatial data. We allow customers to leverage both open-source datasets and their own enterprise data. For example, there are a variety of open-source weather forecast datasets available to the public, and thanks to our fast processing platform, users even without coding skills can generate visualizations that present complex geospatial data in an accessible format.
Our customers span a variety of sectors, including geospatial intelligence, telecommunications, and utilities. Insurance companies also use our products since the analytical insights from this data informs their decision-making when it comes to various risk factors, such as earthquakes and other natural disasters.
Makaryan: How did Heavy.AI develop its machine learning models for analysis?
Flaxman: As an analytics company, we have bet on AI to take our products to the next level. When we initially tried using off-the-shelf open-source AI models, we found that the accuracy was inappropriate for our customers since they are dealing with high-stakes use cases and expect much more accurate insights than companies in other fields may expect. Additionally, our customers expect a high degree of transparency in explaining the processes of how we build our tools and using external AI providers would not work in this case since it would limit the information we could disclose to our customers. Another challenge was that existing models were not well-suited for our particular flavor of structured query language (SQL), which includes more modern spatial and machine learning functions unavailable in the traditional SQL programming language for storing and processing information in a relational database.
So, we decided to create our own models by fine-tuning a large library of open-source models.
We used a combination of industry-standard SQL benchmarks as well as tens of thousands of our custom question-and-answer pairs to imbue the models with a deep understanding of spatial SQL and the specific language, including special jargon, that our target customers usually use. We also incorporated a natural language processing technique called retrieval-augmented generation, which allows users to provide additional context, like formulas or acronyms, to guide the model’s responses. This is crucial for capturing the nuanced, domain-specific knowledge that our customers require.
Makaryan: You recently introduced a new product called HeavyIQ. Can you explain its main features and benefits?
Flaxman: HeavyIQ allows a user to ask a question, and then it generates answers with visualizations, such as maps, charts, and graphs, accompanied by the appropriate SQL query. Such visualizations can be very helpful to our customers not only because they present data in an interesting format, but also because they save time for staff who no longer need to use external visualization tools. HeavyIQ also allows users to customize these visualizations. We are currently working to expand customization options: our goal is to enable users to fine-tune every element of the visualization, from color palettes to axes and scales, through the conversational interface. We currently offer customers 12 input languages for prompts, including French and Spanish. The key innovation behind HeavyIQ is the tight integration between the conversational interface and our custom machine learning models.
This blending of natural language interaction, powerful analytics, and customizable visualization is what we call conversational analytics. The aim is to make exploring complex data and decision-making more intuitive and accessible, especially for users who may not have deep technical expertise.
Makaryan: How do trends in generative AI development impact Heavy.AI?
Flaxman: There are a few key trends we are closely following and incorporating into our product strategy. First, as AI models become prevalent in enterprise settings, there is a growing demand for transparency and accountability. We are focusing on ensuring that we can easily explain how we built the models. Second, connecting the language models to rich, interactive visualizations is a major focus for us. We want to enable users to fluidly explore and refine their analyses through a conversational interface. And lastly, more sophisticated conversational experiences will elevate AI systems’ potential to increase productivity. Rather than just aiming for one-and-done question answering, we are working to create multi-turn dialogues where the models can ask clarifying questions and guide users to the right insights. If you were to go to an expert with a certain question, you would expect them to ask questions and clarify your situation before they give you their expertise, instead of offering a general answer, and this is where I think AI can become a really powerful tool to use geospatial data. As generative AI capabilities continue to advance, our goal is to blend that intelligence with our high-performance spatial analytics to empower our customers with intuitive, reliable, and insightful tools.
Makaryan: What challenges have you faced leading product development at Heavy.AI?
Flaxman: The key challenge for me has been striking the right balance between depth and breadth in our geospatial capabilities. Do we focus on doing 10 things in geography incredibly well, or do we aim for a broader set of 50 or 100 features, even if they are not as polished? This is the kind of fundamental question I need to grapple with as we consider new innovative approaches to delivering data-driven insights.
This ties into the larger challenge of Heavy.AI’s pivot towards generative AI-powered analytics. We have bet on the company becoming a leader in conversational spatial analytics, which means going up against some of the largest tech firms and well-funded startups. At the same time, this transition has been the most interesting and rewarding part of my role. There is very little prior research or literature to draw from when it comes to understanding which geographic information system (GIS) features are the most useful to enterprises. And the rise of generative AI introduces a whole new set of challenges but also exciting new possibilities to upgrade our offerings. Helping shape the future of spatial analytics is a big responsibility, but I am as excited as ever to help redefine how users interact with and derive insights from complex geospatial data.
