Inspiration
Film greenlight decisions combine creative judgment with fragmented market research, comparable titles, audience signals, timing, production constraints, and uncertainty.
Greenlight AI was built around a simple question:
Can an AI agent research a film project like a studio analyst while keeping the final recommendation evidence-backed, auditable, and bounded by deterministic logic?
What it does
Greenlight AI helps producers and studio development teams evaluate whether a film project should move forward.
A user provides a screenplay, treatment, or film concept. The system then runs a multi-stage agentic workflow:
- Project Intelligence — Gemini interprets the creative material and separates supplied facts, inference, and unknowns.
- Research Planning — Gemini dynamically determines which market questions need to be investigated for that specific project.
- Live Market Research — Parallel Search API is called at runtime to retrieve current external evidence.
- Evidence Synthesis — Sources are normalized with provenance, relevance, contradictions, confidence, and evidence gaps.
- Commercial Assessment — The project is evaluated across seven commercial dimensions.
- Deterministic Scoring — Application logic calculates the final score rather than asking the LLM to invent one.
- Decision & Action Plan — The system returns GREENLIGHT, CONDITIONAL GREENLIGHT, REWORK, or PASS with supporting evidence and concrete next actions.
The seven dimensions are:
- Audience Opportunity
- Market Momentum
- Creative Differentiation
- Competitive Position
- Release & Timing Opportunity
- Production Risk
- Evidence Confidence
How we built it
The frontend is built with React, TypeScript, and Vite, with a FastAPI/Python backend.
Gemini runs through Vertex AI on Google Cloud using Google's google-genai SDK. It handles structured project interpretation, research planning, evidence synthesis, commercial assessment, and executive explanation.
Parallel Search API is part of the real production workflow. Research questions generated from each project are sent to Parallel at runtime, and returned sources become the evidence layer used by downstream analysis.
The final score and decision are calculated using deterministic application logic, keeping the language model separate from the final scoring policy.
The application is containerized and deployed on Google Cloud Run.
Challenges
One of the largest challenges was making structured AI analysis reliable enough for an end-to-end agent workflow.
Real Vertex AI requests exposed schema-complexity and latency issues. We reduced wire-schema complexity, introduced stage-specific generation budgets, bounded retries, and maintained full local schema validation.
Another challenge was preventing sourced facts, model interpretation, and recommendations from being blended together. Greenlight AI therefore keeps evidence provenance, contradictions, uncertainty, confidence, and deterministic scoring separate throughout the pipeline.
Accomplishments
The final release successfully completed a real-provider workflow using both Gemini and Parallel Search.
During the verified hosted workflow:
- all structured Gemini stages completed successfully
- 9/9 Parallel searches succeeded
- 44 external sources entered the evidence pipeline
- citation lineage validation passed
- deterministic scoring validation passed
- the hosted end-to-end workflow completed in about 92 seconds
- 83 automated backend, frontend, and browser tests passed
The system has no production mock fallback for Gemini or Parallel.
What we learned
The strongest lesson was that an AI agent becomes more useful when the model is not treated as the source of truth.
Separating research planning, external evidence retrieval, evidence validation, commercial assessment, deterministic scoring, and explanation makes the output significantly easier to trust and audit.
We also learned that confidence should represent the quality and completeness of available evidence, not pretend to predict whether a movie will commercially succeed.
What's next
Future versions could add:
- richer historical entertainment datasets
- configurable studio-specific scoring policies
- portfolio-level project comparison
- collaborative executive review workflows
- additional evidence providers
- regional market intelligence
- continuous monitoring when market conditions change
The core principle will remain the same: agentic research with deterministic decision controls and traceable evidence.
Log in or sign up for Devpost to join the conversation.