Backlot

Inspiration

Film and TV productions waste huge amounts of money building props from scratch that already exist somewhere — sitting unused in another studio's storage or wrap yard. We wanted to fix that: read a script, find what it needs, and match it to real props that already exist.

What it does

Step 1 — The script: A user uploads a script. The app reads it and pulls out every physical prop mentioned — what it is, which scene it's from, and details like era or location. It then checks our internal database of available props and shows matches, with the script line they came from and why they matched.

Step 2 — The agent: If a prop has no match in our database, a Google Agent (built with Google's Agent Development Kit) automatically searches the web for a real-world lead — a similar item from a recently wrapped production, for example. It never contacts anyone or makes a deal itself — it just brings back a lead for a human to review.

How we built it

• Script parsing and prop extraction using Gemini
• A Replit app for the frontend, database, and matching logic
• A separate Google Agent (Python, ADK) deployed to Agent Engine, called by the app whenever a prop has zero matches
• The two systems talk over a simple API call — the app sends prop details, the agent sends back a structured result

Challenges we ran into

• Getting the agent to do real research instead of returning placeholder/fake results
• Connecting a Google Cloud agent to a Replit app — these live in different places, so we had to properly set up the endpoint URL and authentication between them
• Making sure the matching felt smart (matching by meaning and context) rather than just matching exact keywords

What we learned

• How Google's Agent Development Kit and Agent Engine actually work, end to end — building an agent, deploying it, and connecting it to a real app
• The difference between something that looks automated and something that's actually doing real work behind the scenes
• How much of "building an agent" is really about good instructions, not just code

Built With

• google
• replit