Loading the Uri replay…
Loading the Uri replay…
Scene 1 of 4, step 1
Winter Storm Uri · Houston · February 2021
FreezeLine
One message for everyone backfires. FreezeLine decides, house by house, who should drip a faucet, which one, and who shouldn’t.
The city
18,105
homes in ten Houston neighborhoods, each one a real house footprint.
The freeze arrives
Sun, Feb 14, 12 AM
38°F · feels like 29°F · 0 hours below 20°F
0 High · 0 Watch · 18,105 Low
The worst hour · Tue, Feb 16, 6 AM
8,234
homes at High risk of frozen pipes.
12.5°F · 19 hours below 20°F so far
Risk = how hard the freeze is hitting that hour × how exposed the home’s plumbing is (pier-and-beam crawlspace, pipes in outside walls, insulation, outdoor faucets). The same deterministic engine the console runs.
Real Feb 2021 weather (Open-Meteo) · real house locations (OpenStreetMap) Simulated home details
One message for everyone
“Everyone drip.”
At the worst hour, if every home leaves its taps dripping, all 10 areas fall below the 20 psi line where Texas requires a boil-water notice.
FreezeLine’s plan
Drip where it matters.
8,217 of 8,234 High-risk homes drip one chosen faucet. Every area stays above 25 psi.
0
area-hours below 20 psi
with FreezeLine
1,278
of 1,680 if everyone drips
(whole storm)
Stage 2: when people drip more than told
Near Northside
If 74 homes run a second tap, Near Northside drops to 22.7 psi, under the 25 psi floor.
FreezeLine proposes: kitchen tap only at 74 homes, with no one asked to stop → 25.0 psi.
A supervisor approves every Stage 2 alert, even in Auto mode.
Water pressure at Tue, Feb 16, 6 AM, psi
20 psiboil-water line
25 psiFreezeLine floor
Braeswood Place
10.6
0.0
Houston Heights
12.5
0.0
Meyerland
10.6
0.0
Montrose
8.5
0.0
Near Northside
2.5
0.0
River Oaks
14.5
0.0
Riverside Terrace
6.6
0.0
Third Ward
4.5
0.0
Timbergrove
11.5
0.0
Woodland Heights
8.5
0.0
Simplified pressure model: base pressure − freeze demand − drip flow. Calibrated assumptions, not Houston’s pipe network; in a real deployment the utility’s own pressure sensors take its place.
| Area | Everyone drips | FreezeLine plan |
|---|---|---|
| Braeswood Place | 10.6 | 35.4 |
| Houston Heights | 12.5 | 33.9 |
| Meyerland | 10.6 | 35.1 |
| Montrose | 8.5 | 31.3 |
| Near Northside | 2.5 | 25 |
| River Oaks | 14.5 | 39.2 |
| Riverside Terrace | 6.6 | 28.6 |
| Third Ward | 4.5 | 26.8 |
| Timbergrove | 11.5 | 35.9 |
| Woodland Heights | 8.5 | 29.5 |
Outreach · the worst hour
Text first.
Every Watch and High home gets its own plan by text: which tap to drip, or why not to.
Then call.
Homes that don’t reply get a phone call that asks whether they can do the steps themselves.
Then knock.
High-risk homes that don’t answer get a door check.
Crews where needed.
1,729 households said they can’t do the steps. The 50 crews go first to the highest-risk ones, each the nearest free crew in its area, on real streets (median drive 2 min).
1,679 wait in the crew queue. A household only hears “a crew is on the way” once one is assigned.
18,073 texted
5,479 called (no reply to the text)
306 door checks (High risk, no answer)
Street routes are real OpenStreetMap roads; drive times assume 15 mph on icy roads.
How it’s built
Numbers come from plain, tested code. The agent decides what to do with them and explains it; it never computes a risk score.
Data in
Deterministic engine
Agent
Reaching people
Weather
Open-Meteo · real Feb 2021
Homes + streets
OpenStreetMap · 18,105 homes
Sign-ups
/register · consent
Risk engine
weather × plumbing → score
Drip planner
who drips, which faucet
Pressure model
stays ≥ 25 psi · Stage 2
Agent
rules or Claude · explains, acts
Texts
Twilio · TextBee
Calls
voice agent · Whisper · Kokoro
Crews
A* on real streets
Supabase
Postgres · row-level security
City Operations
Next.js · deck.gl · MapLibre
TypeScript · React 19 · Tailwind · h3-js
Python voice agent: FastAPI · Pipecat
Automated tests for the math, the wording and the honesty rule