Hackathon imagines federal agency websites enhanced by artificial intelligence

When he was little, Rohan Chandra didn’t understand what his father, Anupam Chandra, did for a living.

“When [Rohan] was much younger, like maybe seven or eight years old, he came down to my home office in the basement and said, ‘Dad, I always see you in front of a computer. Are you an accountant?’” recalled Anupam, or A.C. 

“I was seven. I didn’t know what he was doing, but he’s never forgotten that,” replied Rohan, or R.C., who has clearly heard this story many times. 

In R.C.’s telling, his father has been coding for longer than he’s been alive, which, naturally, inspired him. 

“I think I started coding at around 14…I would talk to [my dad], like, ‘Look, I made a website or something or I made this block in HTML,’” R.C. said. “So he really fostered my passion for it. And it was really great to be able to talk to him and be able to sound off these ideas because my mom and my brother, you know, they're not technical.”

Father and son competed together in the General Services Administration’s Federal AI Hackathon on Wednesday, when participants were directed to optimize an agency website by developing a way to ensure that large language models create reliable and authoritative information while still being user-friendly. 

Congrats to @ASSYSTInc, the 1st place winner of @USGSA's #AIHackathon! Their innovative solution offers a great option for harnessing #AI for gov services, showcasing creativity, teamwork, & a vision for a smarter future. Thanks to all participants for your contributions! pic.twitter.com/mSfzA81Z5Z

— Robin Carnahan (@GSACarnahan) August 1, 2024

Large language models, such as ChatGPT, are artificial intelligence systems that can uncover patterns in vast troves of texts and can then respond to a user’s query. 

Despite his decades in the field, A.C. was nervous. 

“This is the first time I'm doing a hackathon. I was scared of it. You know, I don't work so well under pressure,” he said. “So I'm leaning on [my son]. He has done hackathons at colleges and stuff like that. So he's my pillar here.” 

R.C. said he and his dad were building a chatbot specific to the Agriculture Department’s website that could enable, say, a farmer to more easily look up recent policy changes. 

Simchah Suveyke-Bogin, USDA’s chief customer experience officer, said she was especially interested in what ideas could come out of the hackathon. 

“It's a huge opportunity to have folks from the outside…to really kind of go rogue, look at different opportunities, look at it from an outward perspective, and just kind of play,” she said. “Those kinds of opportunities don't come often.”

The hackathon — which took place in GSA’s headquarters in Washington, D.C., as well as locations in New York City and Atlanta — is one way the federal government is experimenting with how it could use AI technology to improve its operations.  

“I've been asked a bunch of times ‘what are you trying to accomplish here?’” said David Shive, GSA’s chief information officer, in opening remarks. “You know, the government, we have data. We have citizen- and business-facing products and services. We suspect that those are not architected particularly well to interact with AI. We're looking for you all to help us do some ideation, maybe even a little bit of coding…looking for ways to help citizens and businesses use AI to interact more effectively with their government. The best part about doing these hackathons is there should be some unknown things that we haven’t even dreamt up that you all will.” 

Chris Shaheen — another participant and a captain in the Air Force — previously utilized a partnership between the Air Force and Massachusetts Institute of Technology to learn how AI could be incorporated into his role as a contracting officer. 

“I don't like my job. And the reason why I don't like my job is not because I don't like buying things for the Air Force and getting the customer what they need. I don't like my job because there's so many pain points [awarding a contract]. So what motivates me to be here is anything we can do to streamline that process,” he said. “I want an easy button that says ‘buy this the best way.’ And then we don't have to worry about that. We can worry about why are we buying this? What are we buying next?”

Shaheen, who placed third in the competition, worked to add AI capacity to the Census Bureau’s online datasets. He said it was medium difficulty. 

Our #AIHackathon participants are hard at work, coding solutions and pushing the boundaries of #AI to enhance government services. Watch as these brilliant teams collaborate and create innovative solutions to transform government #tech. pic.twitter.com/laUy43C7qT

— GSA (@USGSA) July 31, 2024

In her opening remarks, GSA Administrator Robin Carnahan urged hackathon participants to remember that the federal government, in particular, has an obligation to balance the speed of AI development with values like security, accessibility and responsibility. For example, while a business creates AI to sell a product, government’s goal is to create AI that can be used by anyone in the country. 

She also told competitors to generate AI that’s “optimized for trust.”

“We're living in a world where there's a lot of confusion about what can be trusted and what can't. Ultimately, it's the government's job to be an authoritative source for data that's important in people's lives. And so we need to maintain that,” she told Government Executive. “When I talk about optimizing for trust, it is just that. We need to know about the provenance of the data that we're using in the large language models that we're applying to make sure it's accurate. We need to be really clear about how our data is used and how it's attributed.”

Popular chatbots, while impressive, are known for sometimes providing inaccurate information. 

For instance, Zach Whitman, GSA’s chief data scientist and chief AI officer, said a large language model might give preference to population data from a random website with potentially inaccurate information over the Census Bureau. 

“Folks on websites will say, you know, I'm a business. I'm in this town. And I'm gonna talk about my town a little bit. And they'll say we have a population of whatever. That type of information can make it into an LLM. And when you're talking about population data, [an LLM] might associate that with census data,” he said. “Like ‘Census does population. I'm going to put these two numbers together.’ And in an LLM context, it might reply that the population of this town is whatever. And they'll actually be sourcing this random website and not necessarily the Census Bureau. But it makes a lot of sense for the LLM to respond in that way.” 

Felipe Millon — who leads government sales at OpenAI, which created ChatGPT and was one of the hackathon’s corporate sponsors — said the event is an opportunity to see how agencies could practically use AI tools. 

“There's a saying that ‘You don't get fit by reading about working out. You have to go to the gym.’ I think it's the same way with AI,” he said. “There's a lot of interest in AI…but the idea is how do we enable government to use AI or learn about AI? And you can't learn about it other than doing it.”

As a citizen, Millon also said he hopes AI can help the government spend money more efficiently and improve public services. 

Along those lines, Shive said in an interview that GSA’s 2019 hackathon offered $25,000 in prize money but generated $780,000 in value because code developed at the event was able to be scaled across the agency. 

He said that anything created at the hackathon becomes government property. The event this year offered four prizes totaling $10,000. 

While participants had to work quickly, Shive encouraged the D.C. competitors to occasionally look out the window, which offered a stunning view of the city. 

“Watch the citizens and visitors of these United States kind of go on about their lives,” he said. “That's the reason we're doing this. We’re doing this so that this government can do its job serving those people better.”