Roomprint
Your home is still in your photos.
After a fire, flood or earthquake, the insurance company asks for a list of everything you owned: what it was,
how old, what it costs to replace. Most families never made one (59% in an NAIC survey; 44% in a Farmers survey).
After the 2025 Los Angeles fires, 57% of surveyed survivors said their insurer wanted every single item listed and
described (United Policyholders, 6-month survey). People write these lists from memory, in a motel, for weeks, and
they forget most of what they had.
But the birthday photos, the Christmas mornings and the first-day-of-school pictures show the rooms behind the people.
Roomprint reads those photos, finds the things in them, prices each one from shop pages, and hands back a list in the
columns the insurer already uses, with proof on every line: the crop of the photo it was seen in, the date of that
photo, and the store page the price came from.
What it does
- Reads each photo. NVIDIA Nemotron 3.5 Super VL looks past the people at the room and lists the things in it,
each with a box on the photo. The crop becomes that line's evidence.
- Keeps the claim clean. Code drops people, pets, food and party decorations (judged by what a line is, so a
child's desk, children's books, a food processor or a dog bed stay on the claim), and moves the house itself
(kitchen cabinets, sinks, counters, flooring) to a separate list: policies pay for those under the building, not the
contents, and an adjuster strikes them from a contents list.
- Counts each thing once. The same TV is in the birthday photo and the Christmas photo. Code proposes a pair only
when two sightings could be the same thing (same room, same kind of thing, no clashing brand, never two from one
photo: a microwave is never compared with a fridge);
Nemotron VL then compares the two crops side by side. A merged item keeps every sighting.
- Prices it from shop pages. Tavily searches for a new like-kind replacement. Code keeps only shop pages (news,
review, blog and publishing sites don't count) and only sentences that name the item next to a dollar amount; Nemotron 3 Super picks the fair ones; code checks every pick (the amount is in
that sentence; not used, refurbished, "was", "starting at", per-month or an accessory; a "set of 4" becomes a unit
price; a price for one piece never stands in for a set, and a set's price never stands in for one thing). Two
stores or more → the median. One store → flagged. None → "needs a price". Books are never priced by the shelf:
the line asks the family for the titles. Nothing is guessed.
- Shows the maths. Age comes from the earliest photo the item appears in. Actual cash value uses the rules State
Farm publishes on its own worksheet (effective age ×1.4 below average, ×0.6 above average, depreciation capped at
80%) with useful lives from United Policyholders' schedule, and each line shows its calculation. California requires
depreciation to be itemized and measurable (10 CCR 2695.9(f)).
- Reminds you of what no photo shows. For each room, the common things not yet on the list (pots and pans,
clothes in the closet, tools in the garage).
- Exports in the insurer's columns. State Farm's Personal Property Customer Worksheet is Xactware's
"XactContents Import Template", the system adjusters price claims in; FEMA's flood contents worksheet
(FF-206-FY-21-106) asks for a web link per item. Both are one click, columns copied verbatim.
- Knows the California clock. Since SB 495 (2026), after a declared disaster and a total loss, the insurer must
offer 60% of the contents limit (up to $350,000) with no list, and can't require one for 100 days. The app shows what
an itemized list adds on top.
Propose, check, or leave it blank
| A model proposes | Code checks |
|---|
| an item and its box | box inside the photo and big enough to show; not a person, pet, food or decoration (a baking sheet of cookies stays, as a baking sheet); fixtures go to the building list |
| "same object as in that photo" | only asked for plausible pairs; never two sightings from one photo; merged only on a confident yes |
| a price | the amount is in a sentence Tavily returned, next to the item's name; new, not "was" or monthly; the same set or single as the line; median of stores |
| an age | from the photo's own date, not from the model |
| the export | totals recomputed; every line carries its photo and its store link |
Models (NVIDIA Nemotron on Nebius Token Factory)
| Job | Model | Why |
|---|
| read a photo: items, boxes, room | Nemotron 3.5 Super VL (reasoning off) | a vision model that returns usable boxes; about 7 s a photo |
| same object in two crops? | Nemotron 3.5 Super VL | a yes/no on two crops side by side |
| pick fair store prices | Nemotron 3 Super | reading many messy shop sentences and judging like-kind |
Reading nine photos and pricing about a hundred things costs about $0.05 in model calls. A daily cap stops calls
before they happen.
How well it works on the sample home
- Reading photos: 72 of 72 clearly visible things found in each of 3 runs over the 9 sample photos, rooms right
9/9, no fixtures and no people leaked into contents (
eval/run_vision.py --runs 3, graded by code against a
hand-made list in eval/ground_truth.json).
- The sample home today: 114 lines; 66 priced (16 from two or more shops, 50 from one shop and flagged);
48 still need a price (5 of them because two shops disagree too much to pick one).
eval/reapply_rules.py
re-applies the current rules to the finished run without reading the photos or searching again (no paid calls).
- Real photographs: 13 stock interior photos (real rooms, not generated; mostly bedrooms, no people in them),
never seen before: 12 read (one answer came back unreadable and is reported), 218 things listed, 114 priced
from shop pages under today's rules (117 before the set and books checks). Checked by hand on two photos: 21 of
about 26 and 21 of about 22 visible things found. Results (names and numbers only, no photos):
eval/results/real-photos-*.json.
- Tests: 113, no network and no models (
pytest -q tests).
Limitations
- The sample family and their photos are generated, and the eval's answer key was labelled on those same photos.
Real camera rolls are messier: low light, clutter, half-visible things.
- Most priced lines rest on one shop and are flagged; 48 of 114 lines need a price from you (the "Your turn" list shows each one and why). The list is a strong
first draft for an adjuster, not a finished claim.
- Two similar things in different photos can still be counted twice when the crops are too small to compare (for
example two flat-screen TVs in one living room). Every line shows its photos so a person can check.
- No family, adjuster or non-profit has used it yet.
Run it
python3.12 -m venv .venv && .venv/bin/pip install -r requirements.txt
export NEBIUS_API_KEY=... # Token Factory
export TAVILY_API_KEY=...
.venv/bin/uvicorn roomprint.web:app --port 8000
.venv/bin/python -m pytest -q tests
The sample family, their home and photos are made up (the photos were generated). Nothing here is legal or insurance
advice; the policy decides.
License
MIT
