Accord

Inspiration

A story can quietly change as it moves between subtitles, captions, audio description, translated dialogue, trailers and social edits.

A subtitle changing “fourteen passengers” to “four” may look small, but it changes what the audience understands. We built Accord to help creative and accessibility teams catch those changes before release.

What it does

Accord compares every version of a film against a reviewer-approved story reference.

In the demo, it finds seven deliberate issues in an animated short: an incorrect number, a missing sound caption, an early identity reveal, overlapping audio description and a subtitle covering an important sign. It also follows the same number error into a recap and trailer.

Reviewers can watch the exact moment, compare the current and proposed caption on the film, edit the wording, timing or placement, approve the repair, verify the result and undo it if needed.

How we built it

Accord is a React and TypeScript web app with a FastAPI and Python backend.

Google Cloud hosts the application. Gemini extracts timecoded story notes from the film, while a Google ADK investigator gathers evidence and suggests bounded repairs. ClickHouse stores versioned track data and lineage, accessed through the official MCP server.

Cloud Run serves the app, Firestore stores project history, and Cloud Storage holds private media. The demo uses an original animated short, The Last Lantern, created specifically for Accord.

Challenges we ran into

The hard part was making AI useful without allowing it to silently change creative work.

We designed Accord so Gemini can investigate and suggest a repair, but a reviewer must approve the story reference and every change. Each repair is tied to the exact version that was reviewed. After a change, Accord checks the updated tracks again before marking them ready.

We also had to make the workflow understandable in a short demo. That led us to add Before and After playback, a simple caption editor and a connected-version view instead of exposing raw technical data.

Accomplishments that we're proud of

• Built a complete review flow from film upload to verified repair and Undo.
• Connected one issue across subtitles, a recap and a trailer.
• Kept the reviewer in control of every applied change.
• Created and tested two original animated scenes rather than relying on third-party footage.
• Verified the main scene through a full 7 → 0 → 7 repair-and-undo run, and a second scene through 2 → 0 → 2.

What we learned

Story consistency is more than translation quality. Timing, accessibility captions, on-screen text and derivative edits all affect what an audience experiences.

We learned that an AI suggestion becomes much more useful when the reviewer can see its evidence, preview it on the real film and adjust it before approval.

What's next for Accord

Next, we want to support team review, richer timeline annotations, more localization formats and stronger reporting for final delivery.

The goal stays simple: every audience should experience the same story.

Instructions to reviewers Open Accord. Enter this reviewer key:

6WJt-7-9BbpcF2a0xW0Y5aQ0ZRlQ7MvFivIUn2_2NFTzI7SFvJxhRzgPvMKFdUw6

This key authorizes project access and changes.

Built With

• clickhouse
• cloud-run
• cloud-storage
• docker
• fastapi
• ffmpeg
• firestore
• gemini
• google-adk
• google-cloud
• mcp
• python
• react
• typescript
• vertex-ai