Inspiration
Casting still feels like a spreadsheet marathon: directors bounce between briefs, headshots, and ad-hoc mood boards just to imagine whether someone fits. We wanted an assistant that keeps pace with the creative process something that turns loose story ideas into a focused shortlist and helps teams visualize the choice before anyone steps on set.
What it does
Cast ingests actor data, lets you search with natural language or a reference photo, and generates AI audition stills tailored to your storyline. The API extracts structured filters from messy queries, runs semantic matches against indexed talent, and returns recommendations plus share-ready imagery that reflects the scene tone and lighting you described. I embedded and indexed around 10,000 actors with thier image for this project.
How we built it
We wrapped the workflow in FastAPI, using Vertex AI for multimodal embeddings (text and image), Gemini for prompt-driven reasoning, and Imagen to render photorealistic audition stills. Actor records are embedded and stored in Elasticsearch for vector search, while the backend orchestrates scene-context prompts, downloads reference assets, and packages results into JSON responses.
Challenges we ran into
Keeping credentials safe while iterating on cloud services took planning, especially around Vertex authentication and API keys. The multimodal embedding endpoints return different shapes for text versus images, so normalizing responses was tricky. We also had to design prompts that describe the scene without over-specifying the actor. Getting a large dataset to work with.
Accomplishments that we're proud of
We delivered an end-to-end casting pipeline: from uploading actor data to querying, ranking, and producing AI-generated stills in one request.
What we learned
Working with Elasticsearch and connecting it to Gemini Api to talk with the indexed data using natural language
What's next for Cast
We plan to add a lightweight frontend for directors, plug in availability calendars from talent agencies, and expand beyond stills into short audition clips. Longer term, we want to introduce bias checks, rights management, and collaboration tools so casting teams can annotate and compare AI-generated options in real time.
Log in or sign up for Devpost to join the conversation.