How a DSI Scholars Team Leveraged AI to Address Real-World Food Insecurity

For more than four decades, the Famine Early Warning Systems Network (FEWS NET) was a vital resource for averting famine and food crises around the world. By providing a centralized repository of climate data, market prices, and conflict trends, it helped both governments and non-profits identify regions at imminent risk of famine. When US foreign aid funding was abruptly frozen in January 2025, this vital resource was suddenly unavailable – at a time when millions were projected to face starvation within months.

The sudden shift was a call to action for Columbia Climate School Professor Michael Puma and his colleagues. 

“A group of us sort of mobilized,” says Puma, whose research focuses on food and water insecurity. “Our first reaction was, ‘Well, let’s get something up and running to fill the gap.’”

Although FEWS NET has since become available again, its suspension inspired Dr. Puma to explore how cutting-edge technologies could help secure access to some of the insights it offers. Through the DSI Scholars Program, Dr. Puma was matched with Columbia MSDS student Shreyas Kamath. Over the course of their semester working together, the team developed an open-source prototype called FamineWatch, which combines an AI agent with a retrieval-augmented generation (RAG) framework to generate key metrics about food insecurity. Because it draws from online resources, FamineWatch is designed be especially helpful to smaller aid organizations that lack the resources to do on-the-ground data collection of their own.

The AI Advantage

Initially, Dr. Puma had imagined building FamineWatch using a fuzzy logic system that might better model the continuum of food insecurity than traditional, threshold-based approaches. Yet Kamath quickly highlighted that the technical maintenance of such a system might be beyond the capacity of the organizations the team hoped to serve.

Instead, Kamath used an agentic AI approach to combine rainfall and vegetation signals with food price data and conflict reporting from Ethiopia to model how shifts in conditions might indicate emerging risk. Crucially, Dr. Puma’s network of contacts in Ethiopia were available to provide the ground-truth necessary for validating the model’s performance.

For Kamath, the project required a new level of rigor compared to his prior modeling work. 

“It’s not just numbers and points on a graph.” says Kamath. With FamineWatch, it is real “human lives that you’re talking about.”

As is often the case for specialized AI systems, getting the right data – and the right level of granularity – was a core challenge. “The biggest problem was just the availability of data,” Kamath says.

Looking Ahead

Despite those difficulties, however, Dr. Puma is confident that advanced technologies like AI—and even quantum computing—can provide important insights for fields like his. For example, Dr. Puma says, “Quantum mechanics are actually quite useful when you’re trying to model human decision-making,” which plays a crucial role in the dynamics of food insecurity. 

“Even someone working on quantum, they’d be like, ‘Oh, there’s no connection with understanding food security.’ Well, that’s not true,” Puma says.

Although quantum computing is still largely on the horizon, agent-based approaches like the one Kamath implemented for FamineWatch can help model complexity that more classical approaches simply can’t capture.

“We’re sitting here trying to build models and trying to predict food insecurity, and it’s just such a complicated pattern,” says Kamath.

For now, the prototype Kamath developed is a valuable proof of concept that has already become a foundational building block for the next phase of FamineWatch: Securing funding to build a fully functioning, open-source tool.

Puma is currently in conversations with global philanthropic foundations to evaluate how the AI-driven approach of FamineWatch could be scaled to support robust, data-driven decision-making for aid organizations at all levels.

“These AI agents [can help organizations] access the data quicker and then run through different intervention scenarios,” Puma says. “There’s a lot of potential there.”