AI Film Cost Optimization Agent

Inspiration

Our team had watched a YouTube video where someone had spent $1000+ making a short 3-minute film, with the majority of the cost stemming from correcting inconsistencies across generated clips and repeatedly generating shots until the correct one was made.

What it does

AI Film Cost Optimization Agent helps filmmakers reduce the real cost and trial-and-error involved in generative AI video production.

Given a filmmaker's shot description and production requirements, the agent analyzes the technical demands of the scene, evaluates viable AI video-generation models, compares their capabilities and costs, eliminates models that cannot satisfy the shot requirements, and recommends the lowest-cost viable production strategy.

Instead of simply choosing the model with the lowest advertised price, the agent considers the real expected cost of producing an acceptable final shot—including duration, model capabilities and limitations, failed generations, and reruns.

Parallel plays a critical role in this decision process. The agent uses Parallel to retrieve current information about AI video models—including pricing, capabilities, limitations, and availability—so recommendations can incorporate current external information rather than relying only on static model data.

The agent also uses Parallel to search for existing stock footage, royalty-free B-roll, and reusable video assets. This allows it to make a more fundamental production decision: sometimes the cheapest way to create a shot is not to generate it with AI at all.

Generation outcomes, reruns, costs, defects, and director acceptance can also be captured as production telemetry, creating a feedback loop between estimated costs and what filmmakers actually experience in production.

Ultimately, the agent is designed to answer one question:

What is the lowest-cost path to an acceptable final shot?

How we built it

We built the application around Google Agent Development Kit (ADK) and Google Gemini, with Parallel as a core partner integration for live production intelligence.

The Google ADK agent follows a multi-step decision workflow. It first decomposes a filmmaker's prompt into technical requirements such as subject, camera movement, motion dynamics, lighting, physics, and character requirements. It then determines which capabilities are essential before evaluating potential production options.

Parallel — Live Production Intelligence

Parallel is integrated directly into the agent's decision workflow and serves two important functions.

First, the agent uses the Parallel Search API at runtime to research current AI video-generation models, including their pricing, capabilities, limitations, and availability. This allows the agent to make decisions using current external information rather than relying only on a static model catalog.

Second, Parallel powers our alternative-footage search. The agent can search for stock footage, royalty-free B-roll, and existing video assets matching the requested shot. This gives the agent the ability to recommend not generating a new AI clip at all when an existing asset provides a more economical production solution.

After gathering this intelligence, our deterministic cost engine normalizes different provider pricing structures into comparable per-second costs, applies expected rerun costs, eliminates technically incompatible models, and ranks the viable alternatives.

Google Gemini provides the reasoning layer that evaluates these tradeoffs and explains the recommended production strategy to the filmmaker.

ClickHouse — Closing the Feedback Loop

We also use ClickHouse as an empirical telemetry layer.

When a filmmaker reviews a generated result, the system can capture the model used, base API cost, rerun count, total realized cost, whether the director accepted the shot, defect categories, and qualitative director feedback.

This creates a feedback loop between estimated cost and realized production cost. Over time, empirical rerun rates and acceptance outcomes can make future cost recommendations more representative of what filmmakers actually spend to obtain a usable shot—not merely what a provider charges for one generation.

The application is deployed on Google Cloud Run, providing a live browser-based filmmaking workflow.

Together, the architecture combines:

Google ADK + Gemini — agent reasoning and workflow orchestration Parallel — live model intelligence and alternative-footage discovery ClickHouse — empirical generation, cost, rerun, defect, and acceptance telemetry Google Cloud Run — deployment of the live application

The result is an agent that doesn't simply compare AI model prices. It researches the market, understands the technical requirements of a shot, considers alternative production methods, estimates the real cost of obtaining a usable result, and learns from production outcomes.

Challenges we ran into

Our biggest challenge was find a problem to work on, as it seemed most of our ideas were already built / were being built by someone else.

Accomplishments that we're proud of

The closed loop agent system for improved video generation model recommendations works, and the code is setup to make it easy to integrate agentic payments if we so choose at a later date.

What we learned

Google is very generous with it's agent token usage via Antigravity.

What's next for AI Film Cost Optimization Agent

Integrating actual agent payments for 3rd party video generation models. That was deliberately left out as we could not integrate AI tooling other than what was allowed by the competition.

Built With

• agent-development-kit
• antigravity
• google-cloud
• google-enterprise
• parallel