AI Property Maintenance Advisor: Recorded Demo v2

AI Property Maintenance Advisor

Photograph the data plate on a piece of equipment and say what is wrong. The tool reads the plate, waits for the owner to confirm the unit before it spends anything more, researches within a fixed budget, and remembers what it has verified so the next question can be answered without searching again. It returns a brief a technician can act on, with every claim cited, or it says plainly that there is no reliable answer.

New in version 2

It remembers what it has verified. A repeat question checks memory before it searches, and each stored fact keeps the exact words from its source. See case 2.
The confirmation is a real stop. Nothing more is spent until the owner confirms the unit and any code, and a paused run can be finished later. See case 1, case 7.
A hard spending cap on every run and a fixed limit on searches and page fetches, both enforced in code; any call the limit blocked is shown in the research trail. See case 1, case 3.
Every step to try first gets a safety check. After the brief is drafted, Jev, a judgment model from TypeSafe, checks each step and can add a safety flag, never remove one, so a flag the writer set always stays. See case 1, where Jev agrees with all three of the writer's flags. On this recording Jev added no flag of its own.
It can use the property's own records. With the unit's service history attached, the brief says when this has happened before and states the unit's age, worked out by code from its install date. Code also works out whether any warranty coverage on file is still active. Shown with synthetic records on a live run. See case 8.

Reading the plate with a confidence for each field, the cited brief, and the refusal to answer work as they did in version 1. This page replaces the version 1 demo, which replayed five cases from the first version of the tool.

In this recording, four of the eight cases end with no reliable answer: cases 3, 4, 5 and 6. Declining is the tool's success state when its sources do not support an answer, so these are honest results: the writing step is told to say so when its sources are not enough or it cannot match the model to documentation, never to fall back on general knowledge of the equipment, and each of these runs says in its own words why what it found was not enough. Cases 3, 4 and 5 ask about real units, so those refusals also count against the tool's own answer bars, as Part 8 of the teardown reports.

Nothing on this page calls an API or loads anything from another site. Each case replays one recorded live run, named by its run ID, and shows the brief file that run wrote, unedited. In those runs Claude read the plate and wrote the brief, Tavily ran the web searches, and Jev, from TypeSafe, checked each step to try first. Every number on the page comes from a live run named next to it, and nothing is estimated.

Pick a case

This page draws each recorded run from its data file with a small script. With scripts turned off, the briefs themselves can still be opened from the briefs/v2 folder.