Seminar: Principled AI for Real-world Impact: Structured Decision-Making under Uncertainty

While heuristics and expert judgment play a central role in real-world decision-making -- from disease testing to infrastructure planning -- these processes often reveal structure that presents opportunities for principledalgorithmic design to complement and strengthen existing approaches. This talk presents two projects frommy postdoctoral research at Harvard, developed through close collaboration with domain experts, thatillustrate how formalizing such problems can lead to effi cient algorithms with provable guarantees.

The first project addresses adaptive disease testing on graphs, where we develop a Gittins index-based policythat is provably optimal on trees and demonstrates strong empirical performance even when structuralassumptions do not hold. In collaboration with the WHO and the Gates Foundation, we are working toward atrial-run deployment of this method in KwaZulu-Natal, South Africa, under the HIV LIFT project.

The second project focuses on health facility planning in Ethiopia, formulated in collaboration with theEthiopian public health institute and Ministry of Health. We model multi-step allocation under online budgetconstraints and region-level fairness requirements as a submodular optimization problem, and designlearning-augmented and greedy algorithms with provable guarantees. We are working on next steps with ourpartners to validate the method and develop an interactive tool for regional planners.

Together, these examples reflect a broader agenda of using theoretical tools to formalize and improvedecision-making in complex, real-world settings by complementing domain expertise with structure, effi ciency,and provable guarantees.

Davin is a postdoctoral fellow at Teamcore, Harvard University. He did his Ph.D. in Computer Science at theNational University of Singapore (NUS), holds a Master's degree from ETH Zürich, and earned twoundergraduate degrees from NUS, in Computer Science and Applied Mathematics. He is broadly interested indesigning practical algorithms with theoretical guarantees and analyzing existing heuristics by providingguarantees under real-world assumptions. During his Ph.D., his work primarily focused on designing andanalyzing algorithms for learning causal and probabilistic models, as well as learning-augmented algorithms.He is currently actively exploring opportunities to apply principled algorithmic techniques to solve real-worldproblems and create a positive social impact.