ux design civic tech responsible ai

Floodline: A Palantir-Style Flood Response Platform with Assisted AI

One operating picture for cutting evacuation decision time during a flood.

year

2026

role

ux designer (self-initiated concept)

sector

public sector / civic tech

decisions made on a partial picture

flood data is scattered. water levels, road closures, vulnerable residents, field-team locations: all in separate systems, arriving faster than one person can piece together. decisions get made late, or on a partial picture.

grounded in UK parliamentary flood resilience reviews and Local Resilience Forum reports. not invented.

a coordinator, deciding under pressure

a council emergency planning officer, working from a control room during a flood. they coordinate multiple agencies at once.

goal

see the whole picture fast, decide evacuation priorities, keep teams aligned

pain

scattered data, data overload, slow phone-based coordination

pressure

lives and property at stake, accountable for the outcome

the journey map found the pressure point

mapped the full response across seven stages: Monitoring → Alert → Investigation → Impact Analysis → Decision → Coordination → Resolution.

Impact Analysis is the emotional low point, the moment the scale of the flood becomes clear. that made it the hero screen.

journey map, seven-stage flood response
journey map, seven-stage flood response

four layers, not just the user

user. operational. organisational. technology. each grounded in the same research.

a systems map traced how sensor data, public feeds, and field reports flow in, how the AI layer processes them, and how one decision, evacuate Zone X, ripples out to agencies, resources, and the public.

needs analysis, four-quadrant grid
needs analysis, four-quadrant grid
systems map, AI-enabled platform
systems map, AI-enabled platform

a dedicated AI panel. clarity over convenience.

three options: a dedicated AI panel (A), AI woven inline (B), AI on demand (C).

chose A. in an emergency, the officer needs to know instantly: fact or machine suggestion. costs some screen space. worth it.

low-fi wireframes, options A, B, C
low-fi wireframes, options A, B, C

Option A: dedicated panel

AI always visible, separated from the data. strength: officer always sees AI input, clear what's AI vs raw data. weakness: split attention between panel and map.

Option B: inline

AI attached to the relevant data. strength: suggestions in context, less hunting. weakness: can clutter, harder to tell AI from fact, risk of over-trusting AI woven into the data.

Option C: on demand

AI hidden until requested. strength: clean, officer stays in control. weakness: in a fast incident, hidden help may not get used when it's needed most.

why A wins

in a high-pressure emergency, the officer needs AI visible but clearly distinct from verified data: that separation is what responsible AI requires, knowing what's a machine suggestion vs confirmed fact.

six screens. one rule: ai suggests, human decides.

monitoring, alert, impact analysis/decision (hero), coordination, resolution, resource management.

dark, map-centric UI. red and amber reserved for severity only. every AI suggestion carries a confidence rating and an approve/edit control. nothing acts on its own.

final screen, monitoring
final screen, monitoring
final screen, alert
final screen, alert
final screen, impact analysis / decision, hero
final screen, impact analysis / decision, hero
final screen, coordination
final screen, coordination
final screen, resolution
final screen, resolution
final screen, resource management
final screen, resource management

the officer stays in control

every AI suggestion: marked as AI, rated for confidence, approved by a human before anything happens.

not a compliance checkbox. a wrong automated call here doesn't mean a bad recommendation, it means people might not evacuate in time.

built to scan, not read

high contrast for control-room conditions. status states that don't rely on colour alone. map labels legible at a glance.

the plan, since this is a concept

no testing run yet. the plan: live-incident task with emergency planning officers: "a flood alert has come in, decide evacuation priorities." measure whether they find and verify AI suggestions, and how long it takes them to decide.

what i learned

the hardest problem wasn't the AI, it was density versus calm legibility under pressure.

next: pressure-test the responsible-AI pattern when someone's actually rushed. and bring Resource Management up to the same fidelity as the hero screen.