Inside the role: pioneering generative AI for business intelligence

Amazon Q is a generative AI–powered assistant that can answer questions, provide summaries, generate content, and securely complete tasks based on data and information in your enterprise systems. Amazon QuickSight is the AWS unified business intelligence (BI) service built for the cloud.

With Amazon Q in QuickSight, customers get a generative AI assistant that allows business analysts to use natural language to build BI dashboards in minutes and easily create visualizations and complex calculations. Business users can also get AI-driven executive summaries of dashboards, ask questions of data beyond what is presented in the dashboards, and get auto-created narratives and presentations highlighting key insights, trends, and drivers.

After getting my undergraduate degree in computer science, I first went into industry as a programmer for a software vendor in banking tech before entering grad school to earn a M.S. and Ph.D. in computer science.

For my Ph.D. work, I switched focus toward machine learning (ML) and worked on applying natural language processing methods to data mining problems. As part of that work, I met Amazon Principal Scientist Andrew Borthwick at an academic conference, and then years later when I graduated, Andrew recruited me to his team at Amazon! I’ve been here almost seven years and have worked on several interesting projects that sit in between engineering and science work.

This is an interesting time. We’ve passed through an inflection point in terms of both technical capabilities (large foundational models) and the broader awareness and enthusiasm for folks to think big in applying this tech. QuickSight is such an exciting place right now, because we have a large, diverse set of customer personas who can benefit from applying AI in useful, thoughtful ways to improve their day-to-day work. Also, I think the fact that BI is the intersection of humans and data-based visual insights makes it a great place to invent with recent advances in multi-modal foundation models, which naturally combine information from language text and visual images.

I think we’re just scratching the surface here in terms of AI capabilities and how humans interact with AI. We finally invented useful chatbots and returned to the past of the text terminal as the primary modality of human computer interaction! As we did 50 years ago, we will surely graduate beyond the text terminal and invent new ways to combine text, visual information, and UI-based patterns for humans to work with AI capabilities.  

Being Amazon, we look for mechanisms, not just good intentions, to accomplish this. We have our Leadership Principles, such as Think Big and Learn and Be Curious, which really capture the spirit of this last year for me. We also have resources and learning mechanisms available to employees to do hands-on learning of new ML/AI tech. I think it’s mostly our Day One mentality where I feel empowered to identify a way to help our customers and build a prototype, write an idea proposal document — which we call a PRFAQ at Amazon — pitch it to leaders, and get buy-in.

We've had the ML-powered Q&A feature in QuickSight since 2020 and were excited by interesting developments like in-context learning and the InstructGPT paper. When the enthusiasm around generative AI began, we felt a bit lucky to have a head start. Early in 2023, myself and Greg Adams, the director of engineering in QuickSight, created a spreadsheet with a dozen+ ideas for potential ways to improve the customer experience with the rapid developments in large language models (LLMs).

It felt like every week we were getting new papers, new models, new breakthroughs. We knew it would be important to be strategic in prioritizing things that had high value, might be feasible, and might be a good fit for rapid prototyping. We knew that to validate our ideas, being a visual project with many non-technical user personas, we would need some prototypes inside the user interface.

A small team formed to begin initial rapid experimentation and tampermonkey-based prototypes. By the summer, we had several independent useful features that internal customers saw value in. This gave us the confidence to turn these into real product features and continue improving them.

We also enjoyed the benefit of leveraging leading foundational models within Bedrock, which isn’t as expensive and complex as re-training a bespoke model setup from scratch. We then launched all our generative AI features in preview at re:Invent 2023, with general availability following a few months later.  

One of the most rewarding parts was the realization, early in the project, that the ability to rapidly prototype was the real game changer in this tech. For much of my last 15+ years building applied ML in software, most real-world production models were bespoke, complex, expensive efforts. That fast feedback through the tools of in-context learning on foundational models really resonated early. Later, as we launched our previews, it was exciting to get customers to try our generative data stories feature. You select some visuals from your dashboard and type a prompt describing the kind of story you want to tell. In seconds, you get a full draft of a data narrative with an outline, formatted sections with charts, and draft content to further refine. Seeing the delight of customers trying this and saying that this will save them time and help them improve their presentations was really rewarding.

In 2023, there were weekly, sometimes daily, developments happening in the world of AI. Sometimes they were important and worthy of attention and sometimes they were distractions. We deliberately kept focus on working backwards from real customer pain points and tried to create mechanisms to share and be aware of meaningful new ideas. It is important to be on a team who shares the same principles and goals, and approaches the work with humility and recognition that this work is hard. The path from the beginning to the end will not be a straight line and that is OK; that is the nature of invention.

QuickSight has a diverse set of customers: some are large enterprises with many core BI teams to build critical, authoritative reports/dashboards, and others are software vendors building their own software and want to provide visuals and natural language Q&A to their own customers. One example of the latter case is Showpad, a platform that provides functionality to sales and marketing teams to optimize buyer interactions. Their customers needed ways to get insights and data points that were not in their dashboards, so they embedded Q in QuickSight into their product, enabling their customers to use Q to drill down into data points. “With Amazon Q in QuickSight, we’re able to give our customers the ability to query data ... we’re already hearing from customers thrilled with this new experience.”

This is probably true for every year for the last 30+ years, but there’s never been a better time to be a working computer scientist! My advice is to remain curious and find the right balance between learning the breadth of new things and specializing deeper on at least some things. There is more to learn in this active field than is possible, so keep exploring to find the parts of computer science that you are really interested in. Once you do, be sure to dive deep into a few valuable areas. Much important intuition to build up is learned in the details and nuances of project work. Amazon offers many great opportunities to learn, and it’s easy to get overwhelmed with information, so ruthlessly prioritize your time making sure that you are able to balance breadth and depth as you begin your career.  

This is an easy one: the most rewarding aspect is getting to work with such a strong team of fellow builders at massive scale. None of the projects I’ve worked on at Amazon have been easy. They have all had large challenges with unclear paths to get to solutions. Thankfully, our culture self-selects for exactly the kind of people you want on your team to work through such problems: people who think big, who are genuinely curious, who work to earn each other's trust, who recognize that invention is hard, and who are humble and know that there will be many twists and turns along the way. When you have a team that works with that perspective, the daunting hard challenges turn into fun adventures.

It’s neat that chatbots have captured so much public imagination about AI and hopefully inspired folks to think big about this technology. One part that excites me is how large foundation models change the calculus of building applied AI solutions.

Current large foundation models are good at following instructions given in natural language and for many use cases can extrapolate how to do a task from just a few well-picked, in-context examples. This makes rapid prototyping and changing model behavior a simpler “prompt engineering” task, which is much faster and more accessible to different skill sets.

Now I ask product managers to experiment in the Bedrock Playground directly to get hands-on experience with the art of the possible. This shift of making sophisticated AI accessible and understandable to a much larger population of builders will change how we build software, even in places that don’t seem chat-related at all!