U'RICH Global Al

• 

  U'RICH GLOBAL

Inspiration

Coding directly against individual, fast-moving vendor SDKs results in incredibly rigid architecture. If a provider changes their contract, encounters a rate-limiting spike, or experiences an outage, the entire multi-modal chain breaks immediately.

What it does

The application hands the execution plan over to the Genblaze SDK.

How we built it

We configured Backblaze B2 Cloud Storage utilizing the native S3-compatible API gateway. This binds directly to Genblaze as our global asset sink, ensuring all large binary objects completely bypass our application's compute memory.

Challenges we ran into

Upstream AI providers are fast-moving targets; a model endpoint that works perfectly at the start of a hackathon can experience a rate-limiting spike or minor payload contract drift hours later. One broken node in an automated storyboard-to-video pipeline would traditionally kill the entire execution graph.

Accomplishments that we're

During testing, when upstream provider endpoints drifted or threw rate-limiting errors, our system didn't crash. The Google Cloud agent architecture dynamically observed the failure, and Genblaze rolled the execution over to secondary models smoothly without dropping user state.

What we learned good

Total agentic freedom inevitably leads to logical drift and wasted tokens. When we first let our supervisor agents loose, they struggled with multi-step execution graphs. By anchoring Gemini 3 using the Model Context Protocol (MCP), we learned that treating system operations as standardized, schema-enforced tools is the best way to get predictable, reliable

What's next for Untitled

We plan to activate Genblaze’s native step caching and OpenTelemetry-driven evaluation loops. By introducing persistent intermediate state cache lookups, the system will instantly skip identical generation requests (e.g., re-running the exact same prompt and seed context) and fulfill the step instantly out of cached B2 storage objects—saving computational budgets and shaving seconds off agent workflow graphs

Built With

• english