Videyo

Inspiration

Generative video projects lose their memory quickly. Characters live in one prompt, story ideas in a document, scene versions in downloads, and production details somewhere else. A creator has to remember how everything connects.

Videyo starts with the story. Its knowledge graph connects characters, locations, ideas, and scenes. Yo, a creative companion attached to the selected part of the graph, can point out relationships that are easy to miss. The same workspace keeps a production record for generated media.

What it does

Videyo turns a video project into a living story graph. A creator can:

• See how characters, locations, motifs, and scenes connect.
• Unwind the graph into story order without losing the original relationships.
• Select any part of the story and discuss it with Yo in context.
• Build a character sheet before shaping the plot.
• Regenerate one scene while preserving the rest of the project.
• Review the provider, model, asset hash, manifest hash, storage location, and verification state for a generation run.

The public app defaults to a credential-free demonstration. It creates deterministic preview media and a verified Genblaze manifest without making a paid provider call. The live production path uses Genblaze to generate an image, store the asset and canonical manifest in a private Backblaze B2 bucket, and return a one-hour signed review URL.

How we built it

The interface is a responsive HTML, CSS, and JavaScript workspace deployed on Vercel. A Python serverless endpoint receives a scene brief and runs one of two paths.

Demonstration mode creates a deterministic SVG, builds a Genblaze run, calculates the asset SHA-256, creates the canonical manifest, and verifies it. This gives judges a safe way to test the full production record.

Live mode uses a Genblaze Pipeline with the OpenAI image provider. Genblaze's ObjectStorageSink sends the generated asset and provenance manifest to a private Backblaze B2 bucket. The endpoint does not call OpenAI or B2 directly. Genblaze owns the provider step, run identity, storage handoff, hashes, and verification.

The repository has automated tests for prompt validation, deterministic asset hashing, manifest verification, and fail-closed live configuration.

Challenges

The hardest boundary was giving judges a working public app without leaving a paid media endpoint open to unlimited requests. Videyo solves this with an explicit server-side live-generation flag and a demonstration mode that exercises the provenance workflow without claiming an external model ran.

Private B2 storage created a second challenge. Generated work should stay private, but reviewers still need to see it. Live runs keep the durable object private and return a signed URL that expires after one hour.

The product challenge was making provenance useful to a creator. Videyo keeps technical evidence near the creative decision instead of turning it into a separate compliance screen.

Accomplishments

Videyo has a graph-to-story interaction that preserves the relationships between story elements while revealing their order. Yo stays attached to the selected node, so comments have visible context. A creator can regenerate one scene and receive a verified production record beside the result.

The working pipeline uses official Genblaze packages. Live output is designed to land in private B2 storage with its canonical manifest. The public demonstration produces deterministic media and passes Genblaze manifest verification. Eight focused tests pass.

What we learned

A generative media tool needs two kinds of memory. Creative memory keeps track of who, where, and why. Production memory records which provider, prompt, model, and asset produced a result. The product becomes more useful when those memories share one interface.

Genblaze makes provenance part of the run itself. B2 gives that run a durable home. This is safer than trying to reconstruct the history after files have already been downloaded and renamed.

What's next

The current hackathon slice generates images for video scenes. Next, Videyo will add image-to-video, narration, and versioned review decisions. The knowledge graph will become a continuity system that can warn when a character, location, or motif drifts between scenes.

Before opening live generation broadly, Videyo will add per-user authentication, rate limits, durable job state, and streamed progress events.

Built With

• codex
• genblaze