Concept

CleanProof is a phone-based, AI-powered proof-of-service and quality-assurance layer for housekeeping and cleaning companies. Cleaners record two short walkthrough videos per job — one on arrival, one on completion — using a lightweight mobile web app. An AI vision pipeline analyzes the footage frame-by-frame, tags each room, checks it against the client's cleaning checklist (counters clear, floors done, bathroom fixtures, beds made, trash emptied, etc.), and compares before/after states. The agent then runs autonomously: it drafts and sends a branded proof-of-service report (photos/clips + checklist scorecard) directly to the client, logs everything for the business owner, and only pings a human supervisor when something needs attention.

How it works (Vision + Autonomous Agent blend)

• Vision and physical-world sensing: the AI reads the actual physical space via video, not documents — identifying rooms, surfaces, and completion state, and flagging pre-existing damage seen in the "before" video vs new damage seen "after."
• Autonomous background agent: no one watches the video live. The agent runs the full pipeline end-to-end — analyze, score, draft report, send to client — on every job automatically, with humans only seeing the dashboard and escalations.
• Escalation logic: if a checklist item is missed, quality score drops below threshold, or new damage is detected, the agent flags the specific timestamp/frame to the cleaner (before they leave the property) and to the owner/manager, instead of silently sending a bad report.

MVP

• Mobile web app: two-tap video capture (before/after) tied to a job ID and address.
• AI backend (vision model API) that segments rooms, matches against a configurable per-client checklist, and produces a quality score plus before/after highlight clips.
• Auto-generated client-facing report (email/SMS with link) sent automatically on job completion.
• Owner dashboard showing job history, quality trends by cleaner, and an exceptions feed (missed items, possible damage) requiring human review.
• Simple escalation alert (push/SMS) to a supervisor when a job falls below the quality threshold.

What it is

Margin is an AI-first contract redlining copilot for small law firms and in-house counsel. Upload a contract, tell it what your firm's standard positions are (or use a built-in playbook), and Margin produces a fully-redlined, track-changes version of the document with clause-by-clause rationale — the first pass a junior associate or paralegal would normally spend hours producing.

Who it's for

• Solo and small-firm transactional lawyers (2–20 attorneys) reviewing NDAs, MSAs, vendor agreements, and leases
• In-house counsel at startups/SMBs with no legal ops team or CLM budget
• Anyone who redlines the "same kind of contract" repeatedly and wants consistency without re-deriving it every time

Why now

LLMs are good enough to reliably classify clauses, compare them to a stated standard, and generate correctly-worded, properly-scoped redlines — the exact mechanical work lawyers currently do by hand in Word. Nobody has made this dead-simple and priced for the long tail of small practices.

How it works

• Upload a contract (.docx or PDF)
• Choose a playbook — a set of preferred positions/rules ("cap liability at 12mo fees," "no auto-renewal beyond 12 months," "mutual indemnification only") — start from firm-provided templates for common contract types
• Margin analyzes the document clause-by-clause, flags deviations from the playbook, and inserts specific suggested redlines with a one-line rationale for each
• Lawyer reviews suggestions in a simple side-by-side UI: accept, reject, or type a chat instruction to adjust wording ("make this cap $50k instead")
• Export a clean .docx with native Word track changes, ready to send back to the counterparty

Business model

Per-seat monthly subscription, tiered by number of contracts reviewed per month; playbook customization included so firms build institutional knowledge into the tool over time.

MVP

• Single supported contract type to start: mutual NDAs (highest volume, most standardized, lowest risk)
• One default playbook (industry-standard NDA positions), editable by the user as plain-language rules
• Upload flow: user uploads a .docx NDA
• AI backend: LLM segments the document into clauses, matches each against the playbook, generates suggested edits + short rationale
• Simple review UI: list of flagged clauses with suggested redline text, accept/reject/edit per item, plus a freeform chat box to request wording changes
• Export button: generates a .docx with native Word tracked changes reflecting all accepted edits
• No accounts/multi-user system yet — single-user tool, manual playbook upload, focused entirely on proving "upload NDA → get usable redline in minutes" before expanding to MSAs, leases, and team/playbook-sharing features