How I built a repair triage system for remote public housing, and why I hide the model's confidence from tenants

Description and use

ReachNT uses an urgency model, H3 hexagon batching, and a CP‑SAT solver to reduce remote repair wait times to 3 days, hiding uncalibrated confidence scores from tenants.

Problem

Remote public housing repairs are delayed because sending a tradesperson to a remote community for a single job is expensive, causing urgent repairs to wait weeks while town repairs are done in days.

Outcome

ReachNT achieved an average cost of $821 per repair, with 9 in 10 urgent remote repairs fixed within 3 days, compared to a cheapest‑first baseline that cost $561 per repair but left urgent remote repairs waiting 62 days.

Workflow

• Collect synthetic repair requests and real geography data
• Train an urgency model on synthetic text (ROC‑AUC 0.84)
• Calibrate model confidence and hide it from tenants
• Batch nearby communities into shared trips using H3 hexagons
• Generate weekly plans with a CP‑SAT solver (median 8 ms solve time)
• Human coordinator signs off each plan
• Provide tenants with plain‑English explanations of wait times

Evidence

The urgency model reads a tenant's plain-language repair report and scores it. ROC-AUC is 0.84 on wording the model never saw during training.

Source

When the model said it was at least 90% sure, it was right 56% of the time.

Source

Across five random synthetic years, shared trips saved $67 to $76 per repair. On a base of roughly $821 per repair, that is a real cut.

Source

Weekly plans come from a CP‑SAT solver. It proved 99.5% of 26,018 weekly plans optimal, with a median solve time of 8 milliseconds.

Source