# Research agenda: trustworthy intelligence for living infrastructure

## Core question

How can computer vision, inexpensive sensors, hydrologic models, and
constraint-based inference produce auditable, privacy-conscious decisions for
small watersheds and aging urban infrastructure when observations are sparse,
noisy, or mutually inconsistent?

## Three dissertation-shaped directions

1. **Tiny-watershed digital twins.** Fuse terrain, rainfall, culvert inventories,
   camera/LiDAR observations, and stream gauges. Enforce conservation and network
   topology as compatibility constraints; quantify uncertainty rather than
   presenting inferred flows as surveyed facts.
2. **Infrastructure inspection under scarce labels.** Extend prior wayside rail,
   weld, and medical-registration work to culverts, drains, retaining structures,
   bridges, or rail assets. Compare higher-order relaxation labeling with modern
   deep models on rare defects and distribution shift.
3. **Socio-technical adaptation.** Join physical risk with econometrics and
   computational social science to study maintenance prioritization, adoption,
   affordability, and distributional effects.

## Field geography

Atlanta supplies urban stormwater, rail, and mixed-jurisdiction case studies.
Sunnyvale–San José–Santa Clara County supplies a hometown technology and water
management network. UC Merced/Yosemite and the broader Sierra provide a mountain
hydrology and climate-gradient counterpart. Any field work requires landowner,
agency, safety, and research approvals; exploratory narratives are hypotheses,
not boundary or flow determinations.

## First-year outputs

- Reproducible benchmark with public hydrologic/infrastructure data
- Methods paper on physics/topology-constrained labeling under missing data
- Open visualization that distinguishes measurement, model, and uncertainty
- Practitioner memo translating accuracy into inspection and maintenance choices