nasl3yn-b2-hackathon

Topic to cinematic video in minutes — script, keyframes, clips, narration, audio, all stored on Backblaze B2 with provenance manifests.

Comment

Inspiration

Video production is slow, expensive, and requires specialized skills. We wanted to prove that an AI pipeline could take a single text prompt and produce a complete cinematic marketing video — script, visuals, voiceover, music — all in minutes, not days. Backblaze B2 provides the durable, affordable storage that makes storing large video assets practical at scale.

What it does

Nasl3yn takes a topic (e.g. "Smartwatch Pro") and runs a fully automated 7-stage pipeline:

  1. Gemini 3.5 Flash generates a scene-by-scene script with English text overlays
  2. Nano Banana 3 Pro creates branded keyframe images per scene
  3. Veo 3.1 Lite produces cinematic video clips from keyframes
  4. Chirp 3 adds text-to-speech narration per scene
  5. Lyria 002 composes ambient audio beds
  6. FFmpeg stitches everything into a final MP4 with transitions and mixed audio
  7. All assets + a SHA-256 provenance manifest are stored on Backblaze B2

Each run is fully traceable — the manifest records which model generated which asset, enabling audit trails and tamper-evident records.

How we built it

Python backend (FastAPI) with a PWA frontend, deployed on Google Cloud Run. The pipeline orchestrates Vertex AI models through Genblaze SDK's Pipeline and StepBuilder classes. Backblaze B2 stores all generated assets and provenance manifests via the b2sdk native API. Models are dynamically resolved from the Vertex AI catalog at runtime — the latest available version is picked each run. The demo reel was assembled from 6 real pipeline runs using ffmpeg.

Challenges we ran into

  • Model reliability: AI-generated text in videos varies in readability across runs — some renders are crisp, others need retry logic
  • Audio sync: Matching narration timing to video clip duration required adaptive padding and speed adjustment per scene
  • B2 integration: The Master Application Key doesn't work with the S3-compatible API — had to switch to b2sdk's native API for proper authentication
  • Provenance at scale: Generating SHA-256 manifests for multi-asset pipeline runs (keyframes, clips, audio, video) while maintaining idempotency on retries

Accomplishments that we're proud of

  • Working production app — nasl3yn.com is live, serving real users who generate marketing videos daily
  • 7-stage autonomous pipeline — from text prompt to finished MP4 with no manual intervention
  • Tamper-evident provenance — every asset is SHA-256 tracked, stored alongside its manifest on B2
  • 5 Vertex AI models orchestrated through a single pipeline with provider-swappable architecture
  • 17 completed pipeline runs producing 78 video clips, demonstrating real-world reliability

What we learned

  • Backblaze B2's predictable pricing makes media-heavy AI workloads economically viable — video clips range from 10-100 MB each and costs stay predictable
  • Genblaze SDK's provider abstraction makes it practical to swap between cloud providers without rewriting orchestration logic
  • Dynamic model resolution (picking the latest version from the catalog at runtime) keeps the pipeline future-proof as models update
  • Provenance manifests are not just nice-to-have — they're essential for trust in AI-generated media

What's next for nasl3yn-b2-hackathon

  • Multi-provider orchestration — integrate GMI Cloud alongside Vertex AI for provider fallback
  • Multilingual support — Bidi-aware text rendering and multi-language narration
  • Custom brand kits — persistent logo, color palette, and voice per client
  • Real-time streaming — WebSocket progress updates during pipeline execution
  • Batch production — generate N videos from a CSV of topics for marketing teams

Built With

  • agentic-ai
  • ai
  • audio-generation
  • b2sdk
  • backblaze-b2
  • cloud-run
  • fastapi
  • ffmpeg
  • firestore
  • gemini
  • genblaze
  • generative-ai
  • google-cloud
  • image-generation
  • llm
  • machine-learning
  • media-pipeline
  • provenance
  • pwa
  • python
  • text-to-speech
  • text-to-video
  • vertex-ai
  • video-generation
Share this project:

Updates