RepairDraft

What it does

A letting agency has dozens of repair cases open at once, and the ones that hurt are the ones that go quiet - a leak reported on Monday that nobody has chased by Friday. RepairDraft reads every open case, decides the one next step (act, ask, wait or escalate) and drafts the messages to the tenant, the landlord and the contractor, so the person running the desk approves rather than writes. The line on the landing page was "No repair goes quiet."

The bit I am proudest of

One Claude call per case does the thinking. Then a rulebook written in code checks every decision, and it can only ever make a decision more cautious, never less. It can turn "act" into "wait" or "escalate", but nothing can turn an escalation back into an action. The model is free to be clever; the code decides how far that cleverness is allowed to travel. On the hackathon's 30-case test it answered all 30, with no fallbacks, in just over three minutes. My own grading of those answers came out at 17 passes, 12 partials, one fail and no safety failures.

What every team got to

27 teams turned up. Afterwards I cut every lettings team's one-minute demo into one public showreel and wrote up the design they had in common. Working separately, we had all built the same shape: hard rules live in code, not in the prompt; the landlord's approval is the control point; an escalation is a handover a person can decide from in a minute, not a failure; and a human stays on anything that moves money or touches safety. My read is that this is what an AI letting agency front office actually looks like today - very good at coordination, still a long way from closing cases on its own.

Where it stands

Built in a day with Isabella Isotta at Dwelly's London AI Hackathon, 10 October 2026, as one of 27 teams. The code stays private; the showreel of every lettings team's demo is public.

Watch the showreel of every lettings team's demo ↗ · Back to Projects