Ingest evidence
Accept PDF, Markdown or text and create a SHA-256-addressed source record.
Backblaze Generative AI Media Hackathon
Turn approved evidence into media people can trust.
EvidenceCast AI is a human-reviewed research-to-media workflow that transforms approved findings into traceable storyboards, narration, infographics and short videos. Genblaze orchestrates media generation, while Backblaze B2 stores the source records, reviews, provider outcomes, media assets and provenance.
The Render Free application may take about one minute to wake after inactivity.
3synthetic judge fixtures
3approved scenes per story
B2durable workflow ledger
100%provider-status disclosure
How it works
A review gate before every media decision
EvidenceCast AI keeps the source claim, reviewer decision and resulting media connected throughout the workflow.
Review claims
Approve, edit or reject evidence cards while source excerpts remain locked.
Plan the story
Build a three-scene storyboard and narration from approved cards only.
Generate and store
Run scene-level Genblaze pipelines and retain requests, progress, outputs and failures in B2.
Verify and deliver
Create subtitles and delivery assets with hashes, manifests and explicit lineage.
Production-minded architecture
Prompt-to-pipeline-to-durable-storage
Each image scene and narration segment is isolated as its own Genblaze pipeline. This supports bounded retries, partial-failure preservation and scene-level regeneration without rebuilding successful assets.
- Streamlit judge-facing workflow
- FastAPI service and fixture endpoints
- Backblaze B2 through S3-compatible storage
- Docker and Render deployment
- FFmpeg subtitles and final-media composition
Trusted source
↓
Human-reviewed evidence
↓
Storyboard + narration
↓
Genblaze scene pipelines
↓
Backblaze B2 ledger
Accuracy principle
provider_generated: false
The completed reviewed fallback delivery uses Pillow and eSpeak and is never presented as a successful provider-generated asset.
Transparent by design
Failures remain part of the record.
Provider requests, progress, model availability, credit failures, timeouts and deterministic fallback status are stored rather than hidden. A provider result is accepted only when the asset, SHA-256 value, B2 provenance manifest and verification checks are complete.
Explore the project