Inspiration

A screenplay can contain production assumptions that go stale quickly: transit schedules, opening hours, weather, seasons, landmarks, access conditions, and other facts that may change before a shoot.

SceneScout turns those assumptions into a bounded research workflow instead of leaving a production team to manually search every detail.

What it does

SceneScout converts a screenplay scene into a current, source-bound production research brief using Gemini and live Parallel Search.

The workflow is:

scene → typed breakdown → bounded questions → live web research → evidence-bound synthesis → explicit status

Gemini first extracts production-relevant factors from the scene. SceneScout then creates at most six focused research questions and sends bounded queries to Parallel Search. Gemini synthesizes the retained research while nominating the exact evidence IDs it used.

Code — not the model — validates those evidence bindings before a research card can receive its final status.

Explicit evidence states

Every card is labeled as one of:

  • CURRENT
  • UNCERTAIN
  • CONFLICTING
  • NEEDS CONFIRMATION

Positive states require valid retained-source bindings. A conflicting result requires multiple retained evidence IDs rather than a narrative claim of disagreement.

If the model invents or references an evidence ID that was not retained, SceneScout rejects that binding instead of silently presenting it as sourced research.

Why Parallel matters

Parallel Search is the live-research layer, not a decorative API call. It supplies current web evidence for the production questions generated from the scene, while Gemini handles structured scene analysis and evidence-bound synthesis.

That separation lets SceneScout use an agent for reasoning without making the agent the source of truth for time-sensitive facts.

Stack

  • Python 3.12+
  • FastAPI
  • Pydantic v2
  • Gemini via google-genai
  • Parallel Search SDK
  • Vertex AI / Google Cloud Run
  • pytest
  • Ruff

Tests run fully offline through injected fakes and SDK-boundary mocks. The bundled demonstration scene is synthetic and original.

Scope

SceneScout is production-research assistance. It does not replace legal, permitting, safety, labor, rights-clearance, or professional location-management advice.

Public source: https://github.com/DannyBaanks/SceneScout

Demo video: https://youtu.be/BLP-MOgWvVE

MIT licensed.

Built With

  • fastapi
  • gemini
  • google-cloud-run
  • google-genai
  • parallel-search-api
  • pydantic
  • pytest
  • python
  • vertex-ai
Share this project:

Updates

Submission history