Rentropy

Rentropy Copilot is an AI copilot for small landlords (1–50 units) that drafts lease-aware tenant messages, checks them for risky wording before approval, and triages maintenance photos into a draft reply and a work order — with a human approving every single item. It never sends, never books, and never touches money. Small landlords rewrite the same high-stakes messages — late rent, deposit replies, maintenance updates — and a badly worded one can turn a routine message into a legal problem. Existing tools target large property managers and lean autonomy-first: the AI sends, the human finds out later. The builder is a landlord himself; this is the AI assistant designed for his own property-management product, Rentropy (rentropy.app, pre-pilot demo). What it does: 1. Lease-aware message drafts (Muse Spark 1.3 via the Meta Model API), generated only from that tenant's record, with a "Facts used" panel showing the record field behind every number. Drafts can be written in English or Spanish with identical figures and checks; reply tone is selectable (Friendly / Firm but fair / Formal). 2. A pre-send risk check, rules first: deterministic North Carolina validators run in code first — late-fee cap (greater of $15 or 5% of rent, G.S. 42-46), deposit clock (30/60 days, G.S. 42-51/52). A validator missing its inputs answers "not determinable," never a guess. A semantic pass then flags threats, coercion, and illegal self-help wording as exact quoted spans, or the flag is dropped. If the review fails, rule-based results are kept and marked partial — a failed check can never look like a pass. 3. Maintenance photo triage: SAM 3.1 reports only what it can see in the photo — no causes, no mold confirmation, no cost estimates — and an unreadable photo produces a request for a clearer one instead of findings. A tenant's report can also arrive as a voice note, transcribed verbatim by Muse Voice Transcribe into the note field (editable, audit-logged) before the same triage runs. 4. Read-only tenant Q&A answered from the one tenant's record only: facts not stored answer "not in the record," and questions about a different tenant are refused before any model call. 5. An append-only audit log: every generation, check, approval, and question writes a row naming the model that ran, and every row expands to the full stored record. Access is RLS-scoped to the demo account; every change routes through the server functions that write the audit rows. Non-negotiables: AI drafts, a human approves per item; nothing auto-sends. One tenant per call — cross-tenant drafts and questions are refused. Every runtime model is a Meta model, all called through one server-side adapter, so no feature depends on a single model surviving a swap. Built during the event (Oct 26 – Nov 4, 2026). Repo: https://github.com/BrodyBuilds/rentropy-copilot — live demo: https://rentropy-copilot-workspace.lovable.app (demo data: two demo tenants, no real users, no revenue).