MuseVault

Reliable AI media operations with automatic recovery and verifiable B2 provenance.

Comment

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for MuseVault

MuseVault

Inspiration

Generative media teams do not have a generation problem; they have an operations problem. A campaign may use several video, image, and audio models, but prompts, retry history, approved outputs, and provenance records often end up scattered across tools. When a provider times out, teams restart manually and lose lineage. When a final asset is approved, they cannot always prove which model, parameters, or source run produced it.

MuseVault turns that fragile process into a reliable production pipeline.

What It Does

MuseVault is an AI media operations workspace for creative teams. A user enters a campaign brief and starts a production run. Genblaze orchestrates the selected media model, applies a fallback chain when a model returns MODEL_ERROR, and preserves the run lineage. A quality policy checks the output before Genblaze's ObjectStorageSink writes the asset and canonical provenance manifest to Backblaze B2.

The workspace makes the result operational: teams can search assets, see quality scores, inspect provider recovery history, verify SHA-256 integrity, apply B2 organization policy, and download the provenance manifest.

MuseVault also treats generation cost as an operational policy. Its default live image route uses the open-source Pollinations API at a $0.00 application budget, while keeping the provider boundary replaceable through Genblaze.

How We Built It

  • Next.js and React for the judge-facing operations workspace.
  • FastAPI for a stateless orchestration service.
  • Genblaze Core for pipelines, fallback models, manifests, and verification.
  • A custom Genblaze SyncProvider adapter for zero-cost Pollinations sana image generation, with optional genblaze-gmicloud support retained.
  • genblaze-s3 for Backblaze B2 storage through ObjectStorageSink.
  • Backblaze B2 hierarchical run paths for assets and sidecar manifests.
  • Optional B2 Object Lock Governance retention for provenance manifests.
  • B2 lifecycle defaults for incomplete multipart uploads and old versions.

How MuseVault Uses Backblaze B2

Backblaze B2 is the durable system of record, not a final export target. Each run writes the generated media and its canonical manifest.json together. The UI exposes the B2 object path, manifest URI, asset checksum, and canonical hash. Hierarchical paths keep campaign assets browsable, lifecycle policy controls intermediates, and optional Object Lock prevents provenance records from being overwritten during the retention window.

The production design uses B2 Event Notifications to trigger indexing and thumbnail work after an object is created, avoiding a polling loop.

How MuseVault Uses Genblaze

Genblaze performs the core product work:

  • One pipeline API across image and video models.
  • fallback_models for model-level recovery.
  • Provider metadata and retry lineage in every run.
  • ObjectStorageSink for B2 transfer and URL rewriting.
  • SHA-256 asset coverage and canonical provenance manifests.
  • Manifest.verify() before an asset is marked production-ready.

Challenges

The main design challenge was making reliability visible. A normal successful generation does not show why an orchestration system matters, so MuseVault includes an evaluation-safe failure drill. It intentionally injects a primary model MODEL_ERROR at the provider boundary. Genblaze then runs its real fallback_models path against the live Pollinations sana route, followed by the B2 write and manifest verification. The image generation, SHA-256 calculation, manifest write, and independent re-download are real and require no paid model credits. This keeps the demonstration repeatable without pretending that an unpredictable third-party outage happened on cue.

Accomplishments

  • A judge-accessible sandbox plus an explicitly labeled live execution path.
  • A live Genblaze service that writes generated media and manifests to B2.
  • Independent re-download and SHA-256 verification of the stored asset.
  • Downloadable provenance records and a clear recovery audit trail.
  • Production-minded B2 policies surfaced directly in the product.

What We Learned

Provenance is only useful when it stays attached to the operational workflow. A manifest hidden in a bucket is technically correct but easy to ignore. Putting verification, recovery, organization, and storage policy next to the asset changes provenance from compliance paperwork into a daily production tool.

What's Next

  • Add B2 Event Notification consumers for thumbnailing and search indexing.
  • Add C2PA signing as an optional adversarial-verification layer.
  • Add human approval roles and campaign-level retention policies.
  • Add cost and latency comparison across Genblaze provider routes.

Built With

Share this project:

Updates

Submission history