Research

Many public decisions spend a fixed budget on the basis of measurements that only approximate the quantity of interest: a city deciding where to act on air pollution, a road operator deciding how fast traffic should move, a health programme deciding which patients to visit. My work treats each of these decisions as a loop of four steps. Observability asks how finely the available instruments can separate one state of the system from another. Attribution asks which cause produced what was observed. Intervention asks how to spend a limited budget given what can and cannot be attributed. Response asks how the system reacts once the intervention is applied. Each of the first four areas below works on particular steps of this loop, and each entry names them.

Air Pollution

Delhi's public monitoring network is too sparse to show everywhere pollution is high, and its readings do not say which sources produced it. This work adds low-cost sensors to find the hotspots the public network misses, reports when a source attribution is supported by the data, and studies how a regulator should decide which sources to audit.

Steps of the loop: observability, attribution and intervention.

Traffic

A short burst of demand can push a road past the density at which it starts to lose capacity, and the jam then outlasts the burst. This work measures how such jams form in three cities and designs speed advisories and admission control that keep road segments below that density without new roadside infrastructure.

Steps of the loop: intervention and response.

Fields from Sparse Sensors

When a sensor network misses the sites that drive the local dynamics, many different fields fit the same readings. FieldFormer models the field where the network actually observes it, and proves a lower bound on the error of any method that tries to reconstruct the whole field.

Steps of the loop: observability.

Public Health

A health programme can act on only a fraction of the patients it sees, so a model helps only if it ranks patients well inside that budget. This work built treatment-adherence models for India's national tuberculosis programme, a lockdown policy robust to uncertain epidemic parameters, and evaluation metrics for medical images whose annotations are imprecise.

Steps of the loop: intervention and observability.

Gene Regulatory Networks

Single-cell sequencing data are sparse and noisy, and the gene regulatory networks inferred from them depend on how the data were processed. This work measures that dependence, reframes the problem as completing a partly known network, and tests whether ranked regulatory edges replicate.

This area is separate from the loop described above.

Page updated

Google Sites

Report abuse