Inspiration

File organization becomes stressful when it is unclear what will move or how to recover from a mistake. FolderUndo started with a question: how can AI suggest useful filing actions while keeping approval and recovery understandable to the person using it?

The goal was to make each proposed move visible before execution and leave a record afterward.

## What it does

FolderUndo is a Linux prototype that currently works with registered synthetic text documents.

NVIDIA Nemotron suggests a filing destination through Nebius Token Factory. The interface shows the source, destination and content hash for review. Explicit approval authorizes that exact plan; a separate Apply action performs the moves locally and records them in a journal.

Undo follows the journal without another model call. If an original path is occupied or a file has changed, the app preserves the files and reports a conflict. Undo reverses recorded moves; it does not restore edited or deleted contents.

A downloadable reference walkthrough lets reviewers try approval, Apply, Undo, conflicts and evidence export without an API key. This reference mode makes no AI calls.

## How we built it

The prototype uses Python 3.12, a browser interface, SQLite journals and Linux filesystem operations.

Nebius Token Factory serves nvidia/Nemotron-3.5-Super-VL for filing proposals. A separate broker handles provider requests and records model identity, token usage and latency. Credentials remain outside the public repository.

Local validation checks proposals before approval. The model cannot authorize file operations or execute shell commands. Exported evidence bundles can be checked offline for internal consistency.

The project includes an MIT-licensed public repository, setup instructions, synthetic development fixtures and a narrated demonstration under three minutes.

Challenges we ran into

The hardest issue was distinguishing a completed file operation from a completed user journey.

In the latest actual NVIDIA run, Apply and Undo restored the original file, but a subsequent automatic scan failed when the bounded filesystem exhausted its file slots. That failure was retained. The current source removes the redundant scan and includes resource-admission repairs. Subsequent reference browser journeys passed, but the full current NVIDIA-and-browser journey has not been re-verified.

Malformed model output also required strict rejection instead of relaxing the expected schema.

Recovery validation remains incomplete: the retained audit covers 1,168 of 1,204 cases, with 36 not run and a controller time-limit overrun recorded. Free live AI access throughout judging has not yet been established.

## Accomplishments that we're proud of

• Demonstrated an actual Nebius-hosted NVIDIA proposal followed by explicit approval, local Apply and Undo. • Verified a clean extracted build through actual workers and browser interactions. • Restored all four original files in a normal reference journey, including their paths, bytes and file identities. • Preserved all original files across five retained files in a deliberate conflict journey. • Verified both browser-downloaded evidence bundles. • Published runnable source, clear instructions and documented limitations while preserving unsuccessful attempts.

These results are bounded synthetic demonstrations, not general accuracy or real-user benefit measurements.

## What we learned

AI suggestions need validation, and people need a clear way to inspect their consequences before acting.

Recovery claims require checking the files themselves. A successful journal entry alone does not establish restoration, and successful restoration does not erase a later interface failure.

We also learned to distinguish provider latency from total task time, reference tests from fresh inference, and local cost estimates from account billing.

## What's next for FolderUndo

The next priorities are completing current-build NVIDIA-and-browser verification, establishing free live access for judges, and investigating the remaining recovery gaps while preserving existing evidence.

The project also needs feedback from intended users and a comparison with manual filing. Expansion beyond registered synthetic copies should follow evidence that the workflow is useful and that its file- handling behavior is sufficiently validated.