Security deposit disputes are lost at move-in, not at move-out

The problem

Independent landlords and small property management firms handle every tenancy turnover the same way. At move-in someone walks the unit, takes photos on a phone, and maybe fills out a paper or PDF checklist. At move-out, often more than a year later, someone else walks the unit and tries to remember what the carpet looked like.

The costs are concrete. Deductions get dropped because nobody can show the stain was new. Itemized statements go out late, and in many jurisdictions a late or vague statement forfeits the right to withhold anything, sometimes with a penalty for bad-faith withholding. Disputes eat hours of back and forth that a small operator does not have.

Existing tools cover the edges. Property management suites track the deposit balance as a ledger entry. Inspection apps produce a nice report per visit. Few of them link a specific move-out finding to the specific move-in photo of the same item, and that link is the evidence that actually settles a dispute.

Constraints

• Budget: the buyer manages a handful to a few hundred units and will not pay enterprise prices or sit through onboarding.
• Skills: inspections are done by owners, leasing agents or a handyman on a phone, often with poor signal in a basement unit.
• Data: photos live in personal camera rolls with inconsistent timestamps, and checklists vary by unit type.
• Regulation: return deadlines, allowed deduction categories and itemization rules vary by state and sometimes by city, and they change.
• Trust: the output may end up in front of a small claims judge, so every claim must trace to dated, unaltered evidence and a human decision.

System design

The core is a deterministic evidence store. Each unit has a list of inspectable items, room by room. Each inspection captures photos per item, records capture time from the device and server, and stores a content hash so a later edit is detectable. Uploads queue locally and sync when the network returns.

A tenancy record holds the move-in inspection, the move-out inspection, the deposit amount and the jurisdiction. A rules table per jurisdiction, maintained by hand with a source citation and review date, drives a deadline clock that starts at move-out and escalates reminders. Deduction arithmetic, including prorating replacement cost by an item's remaining useful life, is plain code against a depreciation table.

An LLM belongs in three narrow places: - Turning the inspector's voice notes into structured item findings. - Pairing a move-out photo with the most likely move-in photo of the same item when the inspector skipped tagging. - Drafting a neutral, factual description of each difference for the statement.

It does not decide whether something is normal wear, set an amount, or interpret the law. Those fields are filled by the operator, and the statement generator refuses to render a deduction without a linked before photo, an after photo and a human-entered reason.

Trade-offs

There is no automated damage detection. Vision models can flag differences, but a confident false positive in a legal document is worse than a slower human review.

The jurisdiction table is manual and will lag new rules. A visible review date and a link to the source keep that risk honest.

It does not replace the operator's management software. It exports a ledger line and a PDF, which is less convenient but avoids a fragile integration on day one.

What I would build first

I would start with a mobile web app for one tenancy: a move-in walkthrough with per-item photos, hashing and offline queueing, then a move-out walkthrough that shows each item side by side with its move-in photo.

Add the deadline clock for a single state and a PDF itemized statement where every line carries both photos and the operator's reason. Skip the LLM in week one. Validate by replaying a few past move-outs an operator already disputed and asking whether the packet would have changed the outcome.

When a decision will be contested later, design the system around linking evidence at capture time, and keep the model out of the judgment it would be asked to defend.