ReachNT
An AI maintenance triage system for remote public housing. It reads tenant maintenance reports and ranks them by urgency, plans shared trips in batches based on an H3 grid, and presents plain-language explanations to tenants rather than model confidence scores.
Timeline (last 90 days)
ReachNT is a demo of a repair triage system built by Team Top Enders for the CDU IT Code Fair 2026 AI Challenge, aimed at remote public housing in the Northern Territory; it reads tenant repair reports, prioritizes by need, batches neighboring communities into shared weekly trips, and provides tenants with plain-language explanations of the wait.
Since no public Northern Territory work-order dataset exists, ReachNT's repair requests are a full year of synthetic data written by the team, but they include 70 communities from real geographic information, housing counts, repair crew hubs, road closures, airstrips and response-time data, and the synthetic requests are explicitly labeled as synthetic within the product.
ReachNT's urgency model achieves a ROC-AUC of 0.84 on unseen phrasing, and as part of an overall system that includes human review it catches 99.3% of dangerous work orders.
When the model claims at least 90% confidence, its actual accuracy is only 56%, making it an extremely poorly calibrated model; therefore ReachNT hides the confidence percentage from tenants, and 42% of dangerous reports are routed to human handling before any decision is made.
ReachNT's batch scheduling is implemented based on Uber's H3 hexagonal grid, dividing the Northern Territory into a hexagonal grid, with communities falling into adjacent cells grouped into shared weekly trips; across five random synthetic years, shared trips save 67 to 76 dollars per repair, against a baseline of about 821 dollars per repair.
The weekly plans come from a CP-SAT solver, with 99.5% of 26,018 weekly plans proven optimal and a median solve time of 8 milliseconds; each weekly plan requires a named coordinator's sign-off before dispatch, and the solver optimizes cost and breakdown days but refuses to sacrifice urgency order.
In one synthetic year, ReachNT averaged 821 dollars per repair, with nine out of ten remote emergency repairs fixed within 3 days and a weighted breakdown days of 61,700; the cheapest-first baseline averaged 561 dollars per repair, but remote emergency repairs waited 62 days versus 3 days in towns, with breakdown days reaching 233,100.
Before release the team conducted a multi-agent AI audit of the system and found a real defect: a dangerous life-support repair could be downgraded to a routine repair based on a single photo alone; the fix has three layers — the portal UI blocks that downgrade path, the server validates, and the database design records the original severity.
The team candidly admits this is a synthetic evaluation, that the repair requests and tenant behavior are fabricated, and that how models trained on synthetic text behave on real reports is unknown; if real work orders were obtained, the top three things would be to retrain the urgency model on real tenant phrasing, rerun the calibration analysis, and test the shared-trip batch scheduling with actual travel patterns rather than scheduled distances.
The team is proposing a pilot in the 16 communities of the Katherine hub.
All Facts (10)
ReachNT is a demo of a repair triage system built by Team Top Enders for the CDU IT Code Fair 2026 AI Challenge, aimed at remote public housing in the Northern Territory; it reads tenant repair reports, prioritizes by need, batches neighboring communities into shared weekly trips, and provides tenants with plain-language explanations of the wait.
50%UnverifiedSince no public Northern Territory work-order dataset exists, ReachNT's repair requests are a full year of synthetic data written by the team, but they include 70 communities from real geographic information, housing counts, repair crew hubs, road closures, airstrips and response-time data, and the synthetic requests are explicitly labeled as synthetic within the product.
50%UnverifiedReachNT's urgency model achieves a ROC-AUC of 0.84 on unseen phrasing, and as part of an overall system that includes human review it catches 99.3% of dangerous work orders.
50%UnverifiedWhen the model claims at least 90% confidence, its actual accuracy is only 56%, making it an extremely poorly calibrated model; therefore ReachNT hides the confidence percentage from tenants, and 42% of dangerous reports are routed to human handling before any decision is made.
50%UnverifiedReachNT's batch scheduling is implemented based on Uber's H3 hexagonal grid, dividing the Northern Territory into a hexagonal grid, with communities falling into adjacent cells grouped into shared weekly trips; across five random synthetic years, shared trips save 67 to 76 dollars per repair, against a baseline of about 821 dollars per repair.
50%UnverifiedThe weekly plans come from a CP-SAT solver, with 99.5% of 26,018 weekly plans proven optimal and a median solve time of 8 milliseconds; each weekly plan requires a named coordinator's sign-off before dispatch, and the solver optimizes cost and breakdown days but refuses to sacrifice urgency order.
50%UnverifiedIn one synthetic year, ReachNT averaged 821 dollars per repair, with nine out of ten remote emergency repairs fixed within 3 days and a weighted breakdown days of 61,700; the cheapest-first baseline averaged 561 dollars per repair, but remote emergency repairs waited 62 days versus 3 days in towns, with breakdown days reaching 233,100.
50%UnverifiedThe team candidly admits this is a synthetic evaluation, that the repair requests and tenant behavior are fabricated, and that how models trained on synthetic text behave on real reports is unknown; if real work orders were obtained, the top three things would be to retrain the urgency model on real tenant phrasing, rerun the calibration analysis, and test the shared-trip batch scheduling with actual travel patterns rather than scheduled distances.
50%UnverifiedThe team is proposing a pilot in the 16 communities of the Katherine hub.
50%UnverifiedBefore release the team conducted a multi-agent AI audit of the system and found a real defect: a dangerous life-support repair could be downgraded to a routine repair based on a single photo alone; the fix has three layers — the portal UI blocks that downgrade path, the server validates, and the database design records the original severity.
50%