Greenlight

A crew of Gemini agents researches your logline on the live web with Parallel Search and writes the greenlight memo a studio exec would — comps, audience, risks, verdict — in about a minute.

Comment

Inspiration

Before anything gets a greenlight, someone has to answer the financier's questions: What are the comps? Who is the audience? What did similar titles make? What's already in development? What could kill this? Development executives spend hours per pitch on that research. Independent writers and producers usually can't afford to do it at all — they pitch blind. We wanted to give every screenwriter the research desk of a studio.

What it does

Paste a logline (plus optional title, format and budget tier) and a development crew goes to work — visibly:

  • Comps Analyst finds 4–6 comparable titles with year, budget, performance and reception.
  • Audience & Market Analyst sizes the audience, reads demand signals, recommends platform, timing, positioning and a budget band.
  • Risk & Competition Analyst hunts for competing projects in development, rights/IP and life-rights issues, sensitivities and production risk.
  • Head of Development reads the three reports and writes a structured, cited greenlight memo with a GREENLIGHT / DEVELOP / PASS verdict and a confidence score.

Every search the agents run, every source they read and every status change streams live into the UI, so the memo is auditable rather than magic. One click copies the memo as Markdown.

How we built it

  • Gemini (google-genai) powers every agent. The three specialists are Gemini function-calling loops whose only tool is search_web; the Head of Development uses Gemini's JSON-schema structured output to produce the memo.
  • Parallel Search API (official parallel-web SDK) is that search_web tool — the crew's only window onto the world. Each agent decides its own objective and 2–5 queries; the crew shares one Parallel session per run. Parallel's excerpt-first results mean a whole run fits in one model context.
  • The three specialists run in parallel (fan-out) and the Head of Development synthesizes (fan-in). Steps stream to the browser over Server-Sent Events.
  • Cloud Run hosts the Node/Express service, built by Cloud Build. Both API keys live in Secret Manager and are injected at runtime; the service identity is granted secretAccessor on exactly those two secrets (Cloud IAM). The image never contains a key.
  • Front end is plain HTML/CSS/JS with Material 3 / Apple HIG sensibilities: one accent colour, quiet surfaces, full dark mode, reduced-motion support.

Challenges we ran into

  • Gemini can't combine tools and structured output in one call, so the crew is two-stage: free-form research, then a schema-bound memo.
  • Models like to keep searching. Declaring the tool but setting functionCallingConfig.mode = NONE on the final round forces them to write findings instead.
  • Keeping the whole run under a minute meant capping search rounds and per-result excerpt length rather than page-scraping.

Accomplishments that we're proud of

In testing, the Risk analyst independently discovered that Apple has an air-traffic-control thriller in development and downgraded a pitch from GREENLIGHT to DEVELOP with a concrete retitle-and-pivot recommendation. That is exactly the call a good exec makes — and it came from live web research, not model memory.

What we learned

Letting each specialist choose its own queries beats templated searches: the agents searched for "in development", "life rights" and "viewing minutes" on their own. Good tools plus a clear role produce better research than clever prompts.

What's next for Greenlight

Coverage-style script upload (full screenplay → memo), a Slack/Docs export for studio teams, talent and comps packaging via Parallel's Task API, and per-studio memory of past decisions in BigQuery.

Built With

Share this project:

Updates

Submission history