Wornpage Projects — Nebius edition

From next step to working tool: plan with NVIDIA Nemotron, build and check in Nebius Token Factory Sandboxes, then review the result.

Comment

Inspiration

A requirement changes after a tool has been built. The tool still works, but does it solve the task you have now? Wornpage Projects keeps the goal, request, tool and evidence together so that useful work survives changes—and people can make an informed decision about what comes next.

What it does

Work organizes a project into tasks, dependencies and completion criteria. Task brings together three views:

  • Brief: explain the purpose, next step and what will prove the task is done.
  • Tool: review a build request, inspect generated source and try the result.
  • Review: compare the request with the result, examine build evidence and record a decision.

The Nebius integration sends an approved task request to NVIDIA Nemotron through Token Factory. Wornpage validates the generated files, Nebius Sandbox runs the model-authored checks, and a bounded repair can address a failed attempt. An isolated browser preview lets the person try the resulting tool. Saving evidence and marking the task complete are separate actions.

When a request changes, Wornpage shows the difference and keeps the previous tool available. Earlier checks remain attached to the request they tested.

The demo: the request changes, useful work survives

The 2:56 video follows a release-review project from its goal and dependencies through a recorded Nemotron-built release-packet checker. It uses real Tab navigation and scrolling throughout the app.

An incomplete packet produces three gaps. A corrected packet produces zero gaps for human review. The checker examines supplied JSON; it does not inspect a live service or approve a release.

Then the request gains a CSV export. The checker still works, but CSV export has not been built. Wornpage exposes that distinction instead of presenting the old checks as proof of the new requirement. The person records CSV as follow-up work, restores the original checker scope and completes that task after reviewing the evidence. The handoff becomes ready next.

The project finishes the walkthrough at 2 of 4 tasks done. Packet definition and the original checker are complete; handoff and CSV work remain open. The total increases from three to four because the CSV requirement becomes its own task.

The recorded checker build used two Nemotron Super calls, one conditional repair and one Nebius Sandbox execution. Its 30 model-authored assertions passed. Independent preview tests separately exercise behavior, keyboard interaction and layout. The video replays this build; filming does not make another model or Sandbox call.

How we built it

Wornpage uses Svelte, TypeScript and Wornpage UI components. Project work is stored in the browser. A Cloudflare Worker serves the app and calls Nebius Token Factory, keeping provider credentials on the server.

Nemotron returns three files: index.html, app.js and checks.cjs. Wornpage validates their structure and supplies the isolated runtime bridge. Nebius Sandbox executes the generated Node checks. The request, source and results stay available for inspection. Once built, a tool responds to its inputs locally in the browser without an inference call for every interaction.

Our approach is prompt engineering, not fine-tuning: task-specific instructions, declared inputs and outputs, source validation, bounded execution and conditional repair. Model-authored checks are complemented by independent tests and human review.

Challenges and lessons

Generating code is only part of the problem. A valid response can contain an incorrect test, and a passing test can miss a requirement. The app preserves failed attempts and separates generated checks, observed behavior and the person's completion decision.

The included Recovery Simulator explores delays, expedited delivery and assembly help. It contains developer corrections to generated code and is separate from the successful checker build shown in the video. Its modeled results do not confirm supplier or staffing availability.

What we added for this hackathon

This edition builds on earlier Wornpage interface and project-workflow ideas. The hackathon work added Nebius Token Factory integration, NVIDIA model routing, planning and revision proposals, task guidance, evidence review, bounded usage controls and the Sandbox task-to-tool workflow. It also added the Work/Task review experience and request-change handling shown in the video. The earlier challenge repository remains separate, and the private Projects backend is not a dependency.

Try it

Open the interactive recorded checker. No sign-in or API key is needed. Load both packet examples, edit the JSON and inspect the results. This is the actual retained tool from the Nemotron build shown in the video; playing it does not make new model or Sandbox calls.

Judges can use the Work and Task demo with the credentials in the submission's private instructions. It supports browser-local projects, task editing, evidence and the bundled examples. Signed-in judges can request fresh Nemotron planning, next-step suggestions, evidence reviews and tool builds. Live use shares an initial 20-model-call allowance; each build permits up to two model calls and two 30-second Sandbox checks, including one conditional repair. Generated results must pass validation and remain subject to human review. Owner saved-run data is not shared. The MIT-licensed source and setup instructions include the Work/Task app and configurable Nebius integration for running your own deployment.

Built With

Share this project:

Updates

Submission history