We will be undergoing planned maintenance on Oct 7th 6:00AM UTC / Oct 7th 2:00AM ET

Rootcheck

Currently agricultural and geo model is lab bound

Comment

Inspiration

What it doe## Inspiration

We wanted to answer a concrete question: can a real, agent-driven pipeline take generative and rendered media, land it in Backblaze B2 with verifiable provenance, and prove every step independently — not just trust a script's own "success" message.

What it does

Rootcheck is a real, working pipeline that produces media (via HyperFrames' headless Puppeteer rendering, and generative image/video via GMI Cloud/Replicate) and lands every asset in B2 through Genblaze, with a real SHA256 hash and canonical provenance manifest for each one. It also includes a soil-cross-section Gaussian-splat viewer pipeline (SPZ → SOG → B2), independently verifiable in a browser with no plugin required.

How we built it

Node/TypeScript handles the splat pipeline and Cloudflare Worker routing (rootcheck-router, a standalone VMFE). Python + Genblaze (genblaze-core, genblaze-gmicloud, genblaze-s3) handles generative-media orchestration and B2 upload via ObjectStorageSink. For non-generative media (a real HyperFrames-rendered MP4, or a screen-recorded live composition), we used Genblaze's ingest_assets() API — the correct, documented API for bringing already-produced bytes into a tracked, provenance-complete manifest, deliberately chosen over Pipeline.step(). A Replicate credential is proxied through a dedicated Cloudflare Worker (replicate-proxy) reading from Cloudflare Secrets Store, so the real API token never touches a client or a local .env.

Challenges we ran into

HyperFrames' automated duration-discovery correctly reported one of our real compositions as near-zero duration — it was scroll-driven, not timeline-authored, and we didn't force a misleading render. Instead we built a second, honest capture path: a real Playwright-scripted scroll-through, screen-recorded live and ingested through the same Genblaze/B2 pipeline. We also hit and fixed a real B2 authentication failure that traced back to a single mistyped character in a copied application key — resolved by testing directly against B2's native b2_authorize_account endpoint to isolate the fault from region/signing red herrings.

Accomplishments that we're proud of

Every claim in this submission is backed by an independent check, not just trusted output: real HTTP HEAD/GET requests confirming B2 assets are live, ffprobe confirming actual video codec/resolution/duration, and manifest reads confirming provenance data matches what was actually uploaded.

What we learned

That "the SDK said success" and "the artifact is actually correct" are two different claims, and treating them as the same thing is where real pipelines quietly go wrong. Verifying independently at every step caught two real bugs (a missing settings.json causing a silent render failure, and the credential typo above) that would otherwise have shipped broken.

What's next for Rootcheck

Wiring the upstream Pencil design-generation step (currently stub-mode server-side) so custom compositions can drive HyperFrames renders end-to-end, and completing the Replicate credential handoff to close the loop on the generative-image half of the pipeline.

Built With

  • gcp.
  • genblaze
Share this project:

Updates

posted an update —

Adding the piece we left out of the write-up: the Gaussian-splat pipeline.

Our core B2 use case is actually a real SPZ → SOG → Backblaze B2 pipeline for Gaussian-splat geology captures — a soil cross-section, viewable directly in a browser with zero plugin, at rootcheck-router.app-f17.workers.dev/rootcheck/soil-cross-section. The splat asset is independently verifiable: HEAD the object directly at s3.us-west-004.backblazeb2.com/RootCheck/rootcheck/soil-cross-section.sog and you'll get a real 200 with the correct content-type — this isn't a placeholder link.

The viewer includes real per-layer hotspot annotations (topsoil / subsoil / bedrock), derived from the exact same coordinate data used to generate the splat geometry, so the annotations can never drift out of sync with the actual scene — visible using the built-in annotation navigator once you're in the viewer.

We also looked directly at Genblaze's own supported-format list and its full issue tracker and found zero prior mentions of splat, point-cloud, SOG, or SPZ formats — as far as we can tell, this may be the first proposal to extend Genblaze's pipeline pattern toward spatial/volumetric media for environmental-monitoring use (a soil cross-section, a vineyard canopy scan, a watershed profile — each becoming a provenance-tracked, B2-hosted spatial artifact the same way Genblaze already tracks provenance for image/video/audio today).

The HyperFrames/Genblaze-ingest pipeline in our write-up is the second, complementary half — this splat pipeline is the first, and it's the part most directly answering "B2 storage for generated media" as a novel media type, not just a novel workflow.

Log in or sign up for Devpost to join the conversation.

Submission history