Clearance Desk

Inspiration

Every screenplay gets a clearance report before it shoots: a legal read that checks whether "Daniel Reyes, attorney, Boston" is a real attorney in Boston, whether the bar in scene 12 is a real bar, whether the song on the radio needs a license. It costs thousands per draft and takes one to three weeks, and it restarts on every rewrite. We wanted the first pass to take minutes and to only re-check what changed.

What it does

Upload a draft. Clearance Desk extracts every clearance-sensitive element (names, businesses, brands, titles, locations, identifiers, organizations), researches each one on the live web through Parallel's Search API, scores it clear / caution / conflict with cited evidence and a clearance-attorney-style recommendation, and for conflicts proposes replacement names that are searched again and verified clean. Upload the revision and unchanged items carry their verdicts forward. Export a full report.

How we built it

• Google ADK SequentialAgent with five stages: extractor (LlmAgent, Gemini 2.5 Pro, strict output schema), researcher (deterministic BaseAgent fanning out to Parallel Search), judge (Gemini 2.5 Flash with structured verdicts), remediator (Gemini proposals re-verified through Parallel), reporter (Gemini 2.5 Pro memo).
• Parallel Search API via the parallel-web SDK, fast mode, 2–3 templated queries per item plus an objective so results are focused on identity collisions.
• FastAPI on Cloud Run, Gemini through Vertex AI, Firestore for job persistence, Secret Manager for the Parallel key, server-sent events to stream progress into the UI.
• The UI renders the screenplay itself with highlights that resolve live as evidence arrives, next to a risk ledger with evidence links and one-click replacements.

Challenges

Extraction needs to capture how the script uses a name, not just the name — otherwise the judge can't tell a namesake from a defamation risk. Letting the model write its own search queries made results drift between runs, which broke revision diffing; templated queries per category fixed it. Balancing cost meant Flash for per-item judgment and Pro only for extraction and the memo.

Accomplishments

A real workflow, end to end, that a producer can use tomorrow: from PDF to a cited report with verified fixes in a few minutes, and a revision loop that only pays for what changed.

What we learned

Deterministic tool stages inside an agent network are what make the output auditable. The LLM belongs where judgment is needed — reading the script, weighing evidence, drafting the memo — not in the middle of the research loop.

The business

Clearance is a mandatory purchase: every distributor requires E&O insurance, and E&O underwriters require a clearance report. Today that report costs $1,000–$5,000+ per feature and is produced by a handful of boutique research shops — none of them automated. We don't replace the attorney who signs the report; we make the research behind it instant. Two products, one pipeline:

• Copilot for clearance houses and studio legal — the shop runs Clearance Desk first, reviews instead of researches, and still signs. Priced per seat/report at roughly $1,000/month per shop; 80 clearance shops, studio and streamer legal-ops teams ≈ $1M ARR.
• Draft-stage self-serve for the long tail — indie features, docs, YouTube and podcast fiction that can't afford a $1,500 report today get a cited first pass at $99 per draft (with a $29/month writers' plan inside screenwriting tools); 15–20k drafts a year ≈ $1.5–2M ARR.

That's a low-single-digit-millions ARR business on the core product alone, before the upside: every brand flagged caution is also a product-placement lead, and the same pipeline extends to chain-of-title and music-rights checks — the rest of the rights-and-clearances stack.

What's next

Pilot with two clearance houses (report export in their house format is the whole integration), chain-of-title and rights lookups for adapted IP, and the product-placement flip — turning brand cautions into a monetizable placement list.

Built With

• cloud-run
• fastapi
• firestore
• gemini-2.5-pro-/-flash
• google-adk
• google-genai
• parallel-search-api-(parallel-web)
• secret-manager
• vertex-ai