Inspiration
In February 2025, a renter near Houston found a house on Craigslist. The poster said they were the landlord and gave the renter a door code. The renter paid $2,200 for rent and a deposit and moved in. Two days later the real owner, a rental company, told them the poster did not own the house (ABC13 Houston, Feb 15 2025).
They did nothing careless. The photos were of a real house, and they even got inside.
This is common:
- Rental scams: the FTC counted nearly 65,000 rental-scam reports since 2020 and about $65 million lost, with a median loss of $1,000. About half of the reports in the 12 months to June 2025 began with a fake Facebook ad (FTC, Dec 22 2025).
- Real estate scams overall: in 2025, 12,368 people lost $275,110,419 (FBI Boston, Aug 11 2026). That covers all real estate scams, not only rentals.
These scams work because real photos pass. The person and the offer are what's fake, and the home's own record can show that. So we built what a renter needs to prepare for a tour, and put a check against the home's record inside it, before any money moves.
What it does
SNAPBACK is rental tour prep with scam safety built in. Find a place, see it on the map, walk it in 3D before you go, and check the ad before any money moves.
Tour prep:
- Listings finder: researched places from CoStar's Houston listings that fit your budget.
- Map: the places on a 3D city map.
- 3D walk: "Walk it first" tours from the building's Matterport scan.
- Tour request: drafted to the leasing office on record. You send it yourself.
Scam safety, built in. Text SNAPBACK an ad before you pay and get a report: is this the real listing, and who do I call? You send the ad on iMessage, or paste it in the one box on the web: a link, its text or screenshots. A team of checks runs. Each reads one source and writes one sourced line on the card:
- Photos taken in this home: the ad's photos are matched against the building's Matterport 3D scan. "Taken in this building (3 of 3 photos)", or "Taken in a different building. Red flag."
- Contact and unit against the record: CoStar's record of the building has the leasing office, its website and its units.
- For sale, posted as a rental: the Homes.com listings in the CoStar kit.
- Owner of record: the county owner roll for Houston houses, shown with that owner's own published rule on where it takes applications.
- Public records: the Texas Comptroller's company check, HUD's manager record and Houston address points. Each is a neutral information line that names its source.
- The ad's own words: asks for money, papers or haste. "Zelle before a viewing. Red flag."
- Here now: the poster is asked for a fresh photo of a spot in the home, and a copy of a listing photo is caught.
The card gives one of four verdicts:
- Hold off for now
- Don't pay. Something here doesn't match.
- Could not place these photos
- Checks passed. Still follow the two rules.
Code sets the verdict. "Don't pay" needs proof, and SNAPBACK never calls a named person a scammer.
In iMessage, it's a conversation, not a form:
- Photos get an instant "Got 3 photos. Checking where they were taken, about 20 seconds." Then comes a short report: the verdict, the top two reasons, one next step and the full report.
- "Why?", "Who do I call?", "Is it safe?" and "How did you check?" are answered from the same case. "How did you check?" gives one line per agent. None of them starts a new check.
- It remembers the renter's language, budget and beds for 30 days. "Show me safer homes" uses them, and "forget me" deletes them.
- When a listing is being copied, it offers to draft a note to the real leasing office. The note is marked "Draft, not sent. You send it."
- A voice memo in gets a spoken reply out of 15 seconds or less: the verdict, the top reason and who to call. It never speaks a phone number, a rent or a link.
On the web:
- The report shows the ad beside the real listing and a hand-off link to it.
- The Proof Room lets you watch the check team work: a Planner, Rules, Record, Phone and office, Owner and licence, Photo and Web, with the Decider (code) as the only one that sets the verdict.
SNAPBACK never pays, books or sends anything for you.
What a realtor said:
"if he can add in additional safety features for preventing further scams and financial scams, I think this would be a very good app to have in the marketplace."
Ashwin Veeravalli, Realtor, The Agency, Los Gatos CA
How we built it
- Backend: Python 3.12 and FastAPI. Intake, the rules, the case runner and the API are in
src/snapback/. - Code sets every verdict.
verdict.decide()is the only place a verdict is chosen, and no model ever sees or sets one. - LangChain deepagents runs on every case, in the background. A planner writes todos and hands each check family to its own sub-agent. Each sub-agent has one read-only tool and writes its findings as notes in the agent file system. The answer never waits on the agent, and a hard time limit stops it.
- Photo placement: DINOv2 retrieval over views rendered from the Matterport scans, then DISK and LightGlue matching (kornia), RANSAC verification and a spatial-spread check.
- Data:
- The CoStar kit (property records, Homes.com listings) and Matterport scans. We were also given Vexcel aerial imagery, which is shown as a picture and not yet used in a verdict.
- Harris County owner records, TREC licence data, the Texas Comptroller, HUD and City of Houston.
- The kit stays in a git-ignored folder. Only derived yes/no facts reach a model, never kit rows, photos, addresses or prices.
- Intake: local OCR (RapidOCR on ONNX Runtime) for screenshots, listing-field extraction and a prompt-injection guard.
- iMessage: a Node sidecar on Photon 2.0 (the Spectrum SDK).
- Only allowlisted conversations are answered, and their ids are kept as keyed hashes.
- It never texts first.
- Voice uses ElevenLabs speech-to-text and text-to-speech through our API, sent as a native voice bubble.
- Front end: Vite, React and TypeScript, with a MapLibre 3D city map, an isometric house, a walkthrough tour and the check team's crew view. Every screen says LIVE or FIXTURE.
- MCP server: SNAPBACK as tools for any agent.
- Forensics (Arlequin): a ring finder that links ads by reworded text, shared photos, numbers and wordings, with Louvain communities. Every finding carries its evidence.
- Process: four people, each running coding agents in their own lane.
- Every change went through a PR against shared contracts, with failing-test-first fixes and review.
- CI ran on every change: a hygiene guard, ruff, ESLint, Prettier, TypeScript, pytest, the web and sidecar tests, and an eval gate.
Challenges we ran into
- Not overclaiming. Real photos of a real home pass any photo check, so the product has to separate "the photos are real" from "this person may rent it". We kept "Don't pay" for proof only. We quoted no accuracy number until we'd measured it, and we label each number with its n, and with renders or photos.
- Privacy in public records. The county owner roll names people. We changed what the card shows: a private owner is "a private owner", and a company is named only when it matches one we already know. Each privacy gate is tested on its own by turning it off.
- Sponsor data rules. No kit data in git or in a log, and only derived yes/no facts reach a model. A hygiene guard in CI holds the line.
- Messaging is messy. iMessage photos arrive as grouped messages. Voice memos come in Apple's CAF format and have to be converted. A spoken reply must never read out a phone number. And a reconnecting event stream must never text the same alert twice.
Accomplishments that we're proud of
- One rule held from start to finish: code sets the verdict, and every line on the card says where its fact came from.
- Photo placement on real listing photos, LIVE (CoStar OK'd sharing the totals): "Measured on 78 real listing photos: 43 of 48 placed in the right home, 0 in the wrong one, 0 of 30 decoys fooled it." The right room is not measured, and we don't claim it.
- A live tamper test (LIVE; a real listing and tampered copies of it, nothing posted anywhere):
- a changed phone number, caught 6 of 6 on real photos and 2 of 2 on scan renders;
- "send the deposit by Zelle before the tour", caught 6 of 6 and 2 of 2;
- a lookalike website link, caught 6 of 6 and 2 of 2.
- Batch run (FIXTURE, composed ads): 117 of 123 handled as expected, including 43 of 43 clean messages left alone.
- The news case we started from replays end to end and stops at Hold off before the $2,200 goes out (FIXTURE replay).
- Speed (LIVE): text-only checks take about 1 s, and checks with photos about 19 to 26 s.
- The spoken iMessage reply is 11.8 to 12.5 s long, and making it takes 1.31 to 1.55 s (ElevenLabs, n=3, on the demo laptop).
- Forensics (SYNTHETIC, n = 301 ads with 4 seeded rings):
- 4 of 4 rings recovered and 0 of 244 honest ads flagged; a keyword baseline found 0 of 4.
- On 10 held-out seeds, ring precision was 0.88 and recall 0.93.
- The corpus and the method were built together, so this is not an accuracy claim for real listings.
What we learned
- The strongest signal is often not the photo but the record behind it: who owns the home, who manages it, and whether it is for sale.
- Saying "not checked" honestly builds more trust than a confident guess.
- Renters won't download an app in the middle of a scam. The check has to live where the ad arrives: a text thread.
What's next for Snapback
- Edited and AI-made photos. One AI-edited photo was caught as inconsistent with the real scan (n=1). A stricter bar now runs, but at n=1 it needs measuring before we claim more.
- University partnerships for international students, who rent from abroad, sight unseen and to a deadline. Pre-arrival housing guides could carry "check any listing before you send a deposit".
- Platforms: CoStar could ship this beside its listings, and other platforms could run the check when an ad is posted.
- Reach: more cities and languages (English and Spanish today), and accuracy measured on real ad photos at scale.
