Inspiration
Every year, millions of residential tenants and small commercial lessees lose substantial portions of their hard-earned security deposits to unfair, inflated, or pre-existing damage charges.
Tenants often possess the proof needed to dispute these deductions—move-in walkthrough photos, signed lease agreements, email maintenance requests, and hallway security camera logs. However, this evidence is scattered across terabytes of personal storage, difficult to organize, and far too sensitive to upload to public third-party cloud AI models.
LeaseDiff was inspired by a simple premise: your private home data belongs on your own hardware. By leveraging the local compute, native multimodal models, and private storage of the UGREEN AI NAS (UGOS), we built a tool that acts as an objective, cryptographic forensic auditor—turning a stressful landlord dispute into a mathematically verified rebuttal packet without a single byte leaving the user's home network.
What It Does
LeaseDiff (lease-condition-diff) cross-examines a landlord’s itemized move-out deduction invoice against years of local multimodal evidence to generate a hash-verified deposit rebuttal dossier.
Instead of a generic file organizer, LeaseDiff executes a 4-Way Temporal State Diff across four timeline anchors:
- Move-In Baseline ((T_0)): Inspects original move-in photos and checklists to prove pre-existing wear.
- Tenancy Maintenance ((T_{\text{notice}})): Cross-corroborates unresolved work orders and contractor invoices.
- Key Handover ((T_{\text{vacate}})): Verifies final unit condition and door sensor/surveillance entry logs between departure and landlord inspection.
- Landlord Claim ((T_1)): Scans the itemized invoice and reconciles claims against lease wear-and-tear exemption clauses.
Deterministic Verdict Engine
Every line-item charge is assigned one of six deterministic verdicts with mathematical depreciation formulas: $$\text{Liability} = \max\left(0, \text{Claimed Amount} \times \left(1 - \frac{t_{\text{elapsed}}}{L_{\text{useful}}}\right)\right)$$
PRE_EXISTING_AT_MOVE_IN: Photo EXIF and visual feature pairing prove prior damage ($0 tenant liability).REPORTED_UNREPAIRED_DEFECT: Structural or plumbing issues previously reported to maintenance ($0 liability).POST_VACATE_CHAIN_BROKEN: Doorway camera timestamps confirm third-party access after key handover ($0 liability).CLAUSE_EXEMPT_OR_PRORATED: Normal wear-and-tear or prorated based on contractual carpet/paint useful life.VALID_OR_UNREFUTED_CLAIM: Legitimate tenant responsibility (honest, fair assessment).NEEDS_MANUAL_REVIEW: Blurry photos or stripped timestamps flagged transparently for user review.
The output is a complete rebuttal package: Deposit_Rebuttal_Dossier.md, Line_Item_Audit.csv, and an Evidence_Manifest.md pairing each charge with SHA-256 fingerprinted exhibits.
How We Built It
LeaseDiff was built exclusively to the official UGREEN Uliya Skill Development Standard (uliya-skill-kit v0.7.3 / Capability Catalog v2.0.0) as a pure Markdown parent orchestrator with modular reference specifications:
file-agent(ug-filesystem,ug-filemgr): Enforces strict read-only access on source directories, calculates SHA-256 fingerprints, deduplicates redundant scans viascan-dedupe, and manages exhibits.document-agent(ug-document): Employscontract,invoice, andreceiptextraction schemas to parse legal clauses, statutory deposit return deadlines, and itemized billing rows.photo-agent(ug-photo): Runs local scene descriptions, EXIF extraction, andvision-ocrto match architectural features between (T_0) and (T_1).security-monitor-agent(ug-surveillance): Analyzes entryway event intervals between key handover and landlord inspection without invasive facial recognition.knowledge-agent(ug-knowledge): Transcribes walkthrough voice memos and writes the structured dossier and CSV audit sheets.
Challenges We Faced
- The Non-Guessing Principle: In legal and financial disputes, AI hallucinations destroy credibility. We enforced a strict evidence-grading hierarchy: timestamps degrade from
EXIF_VERIFIEDtoCROSS_CORROBORATEDtoINFERRED_TIME_ONLY. If a photo lacks verifiable metadata, the engine refuses to guess and downgrades confidence. - Handling Stripped EXIF & Device Clock Skew: Photos shared via WhatsApp or chat apps often lose EXIF data. We designed a fallback pipeline that searches the frame for dated move-in checklists or newspapers using
vision-ocrto reconstruct capture time. - Read-Only Safety & Approval Gates: A user’s private NAS contains irreplaceable personal memories. We implemented strict architectural barriers ensuring zero files are deleted or moved, and no outputs or tags are generated without explicit dry-run confirmation through the UGOS approval gateway.
Accomplishments That We're Proud Of
- Zero External Dependencies: 0 third-party binaries, 0 scripts, 0 external cloud API keys. The entire orchestrator runs natively inside the UGOS ecosystem.
- True Multi-Agent Collaboration: All 5 registered Uliya agents (
file,document,photo,surveillance,knowledge) work in deterministic harmony. - 100% Privacy Preservation: Tenants can process 50+ gigabytes of 4K photos and leases without exposing intimate home interiors to cloud training pipelines.
What We Learned
- How to architect complex temporal state machines using pure declarative orchestration.
- The power of on-premise AI NAS: local NPU/GPU inference eliminates upload bottlenecks for large media collections.
- Best practices in prompt-injection shielding when handling untrusted user documents and landlord claims.
What's Next for LeaseDiff
- Round 2 Hardware Testing: Deploying onto the physical UGREEN AI NAS iDX6011 Pro to benchmark on-device model latency and high-volume parallel photo batching.
- Multi-Jurisdiction Tenancy Law Templates: Integrating state- and country-specific statutory deposit return timelines and statutory penalty calculation routines.
