1
Detect what is starting now
Every 311 report on the map, a live pull of today's reports, and a subset-scan alarm for clusters growing faster than their own history.
Built for HackUTD 2026, CBRE track: Predict Before It Spreads
Property distress spreads like fire. See where it goes next, and where to stop it.
Firebreak forecasts every building in a city for the next 90 days, traces each new problem to the building that set it off, and plans the few repairs that spare the most neighbours. It runs on public records that share no IDs, and it stops at a person's approval.
- Buildings scored
- —
- City reports joined
- —
- Top 2,000 that got a report, held-out quarter
- —
- Median warning before the first call
- —
01 · The problem
One neglected building raises the odds for every building around it.
A code violation, a vacant house, a landlord who stops answering. Each one is logged somewhere: 311, the appraisal roll, a permit system. None of those systems share an ID, so no one sees the pattern until the block has already tipped.
Dashboards show what already happened. Inspectors still work the queue in the order calls arrive. By the time a cluster is obvious, the cheapest fix, at the building where it started, is months behind.
—higher odds of a new report when the building next door had one in the last 30 days
—higher odds when another building with the same owner had one: a hidden dependency no map shows
—of Dallas property reports were most likely set off by an earlier report nearby (Hawkes model)
—for shared pipes and streets. We tested it, and it isn't a real link. We say so.
02 · How it works
Five screens, one loop: detect, predict, simulate, act, and prove it.
2
Predict the next 90 days
Gradient-boosted risk for every building, calibrated with Venn-Abers intervals, plus a certified list that promises at most 10% wrong, and a neural field for the blocks nobody calls about.
3
Simulate one building, 400 futures
Click any building. A spatio-temporal Hawkes model plays out its cascade 400 times, then a CELF search finds the three repairs that spare the most reports at other houses.
4
Act: from alarm to approved case
Press R on a cluster. The playbook traces it to its origin, simulates it, costs the work orders and routes a crew, all with the console's own tools, then waits for a person to hold Approve.
5
Evidence for every claim
A model card with each test on a quarter the model never saw, a fairness audit by council district, and the experiments that failed.
03 · In use
A Monday with Firebreak
A code-compliance supervisor in Dallas, start to finish. Every step is a real screen in the console.
- Home
The brief names the strongest cluster this week: — new cases in — days where — were expected, across — buildings.
- Resolve
The playbook traces the cluster to the building where it most likely started, runs 400 futures and picks the firebreaks.
- Approve
The case shows cost, reports prevented and a 2-opt crew route. The supervisor holds Approve. Nothing is sent anywhere.
- Dispatch
The route opens in Google Maps for the crew; the case moves to Crew dispatched with a timestamp.
- Resolve
The crew reports back. The case closes with an outcome, and the export goes into the department's own system as CSV.
04 · For teams
One engine, three jobs. Each team gets its own Home.
Code compliance
City inspectors, neighbourhood services, council offices
- Morning brief with the clusters to work first
- One-page council district brief and tomorrow's route
- Budget plan ranked by reports prevented per dollar, never by property value
- District-fair re-ranking, audited with the four-fifths rule
— cases a 50-visit-a-day crew reaches before the first call
Owners and managers
Portfolio, asset and property managers
- The share of a portfolio's risk that comes from buildings it does not own
- Which outside buildings to fix or buy first
- A stress test: inject reports at any building and watch the cascade
- A watchlist that flags new 311 reports at or near your buildings
— business tenants linked to the buildings they occupy
Acquisitions
Investors and wholesale buyers
- Distressed, non-owner-occupied leads, scored for expected fee per contact
- Offer waterfall and seller odds by discount band
- Natural buyers: the portfolios next door that lose most if a house tips
- Drive-for-dollars loops, routed and handed to Maps
— leads today, owner-occupied homes never included
05 · Proof
Tested on a quarter the model never saw. Including where it lost.
The model was frozen on — and scored on the 90 days that followed. Every number on this page is read from the same files the console loads.
| Claim | Result |
|---|
06 · Platform
Built to move to the next city.
The pipeline needs two inputs: a property roll with locations, which every county appraisal district publishes, and a time-stamped event log with coordinates. Dallas 311 comes from Socrata, and many US cities publish the same kind of log through the Open311 GeoReport v2 standard. Swap the two loaders and every screen, test and audit reruns.
The output is a set of static files with a written contract, so it serves from any CDN with no database, and the console renders — buildings in the browser.
| Endpoint | What it holds |
|---|---|
/data/meta.json | Counts, the held-out test, edge ratios, assumptions with sources |
/data/buildings.bin | Columnar building table: location, class, risk, Distress Factor inputs |
/data/events.bin | Every 311 report, sorted by day, placed on a building |
/data/hawkes.json | Spread model: parameters, branching, spread rate by month |
/data/alarms.json | Cluster alarms and their backtest |
/data/falcon.json | Neural field, training log and head-to-head backtest |
$ curl -s https://<host>/data/meta.json | jq '.model.network'
{ "auc": —, "watch_list": 2000,
"watch_precision": —, "base_rate": — }- TodayStatic files, CSV and GeoJSON exports, live 311 polling, cases kept in the browser
- NextA shared case store and sign-in per team, alert subscriptions, an Open311 loader for a second city
- NextInside one campus: assets as nodes, work orders and alarms as events
07 · See it run
The promo is the product.
Film mode plays the guided tour full screen, between a title card and an end card. Every shot is the live console. Press F in the console, start a screen recording, and stop when the end card fades.
Browsers only allow full screen after a click, so a direct link plays in the window. Press F inside the console for full screen.
08 · Pricing
Priced to grow with the ground you cover.
Cities and portfolios pay per building watched; buyers pay per seat. Illustrative pilot pricing for discussion, not a quote.
City
$0.10 per building a year
Dallas, — a year: about the cost of — inspections at the console's placeholder rate.
- All five screens for every department
- Council district briefs and fairness audit
- Unlimited users, cases and exports
Portfolio
$2 per building a month
Your buildings and everything within 700 m of them, scored and simulated.
- Neighbour exposure and buy-or-fix list
- Watchlist flags and stress tests
- Portfolio CSV and GeoJSON exports
Acquisitions
$149 per seat a month
Lead lists refreshed with every build, scored by expected fee per contact.
- Deal desk and buyer teasers
- Natural buyers next door
- Drive-for-dollars routes
09 · Questions
What people ask first
Does it send anything to anyone?
No. There is no email, SMS, mail or call integration. Every plan ends at a person's approval, and cases are exported, not sent.
Does it name owners or residents?
No. Owners are hashed before anything reaches the browser, homes are shown at the 100-block level, and owner-occupied homes are never on a lead list.
Is the playbook an AI agent?
It is an agent in the plain sense: a fixed plan of the console's own tool calls, each reading the last one's result. It is not a language model, so every number in its log is computed, and it stops for a human.
Does it rank rich neighbourhoods first?
Never by property value. The budget plan ranks by reports prevented per dollar; ranking by value sent the whole budget to the wealthiest district, so we dropped it.
Where does it fall short?
311 measures what people report, not what exists. Our neural network lost to the boosted model, and the case-closure study found no causal effect. The Evidence screen shows all three.
How fresh is the data?
The Home screen's data-health table shows the age of every source. The live feed polls Dallas Open Data from the browser and re-scores buildings near each new report.