This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady. Dock is a small place to leave unfinished work before a break. Paste a few scratch notes, click once, and get a return card: one note to start with, what is already finished, what depends on something else, and the context worth keeping.
The audience is someone switching between small projects who wants to stop carrying the whole task in their head. The app asks for rough notes, not a perfectly written handoff.
Its deliberate constraint is that the AI cannot write the card's note text. Qwen2.5 7B chooses categories and a note ID. Dock displays the original wording from the input. The card can be corrected and explicitly saved in the browser, then reopened after the current session is cleared.
[Watch or download the 63-second demo](https://github.com/C12isme945/dock/releases/download/v0.1.0/Dock_demo.mp4)
[Public demo release and source download](https://github.com/C12isme945/dock/releases/tag/v0.1.0)
The video is a captioned sequence of actual browser captures, edited for length. It uses invented garden notes and a real local-model run; it is not presented as continuous real-time footage.
The sequence is input → local inference → choose a different return step → save → clear → reopen. The final card restores the chosen note, “Next: sketch where the three pots will fit,” with its original N6 reference.
The application was started in an empty directory on 24 October 2026. It is a new project for this challenge. Qwen2.5 and Ollama are existing dependencies, credited in the README; their weights and runtime are not claimed as new work. Application implementation, documentation and this write-up were AI-assisted.
The inference path is intentionally short:
The frontend is plain HTML, CSS and JavaScript. A Python standard-library server binds only to the loopback address and calls local Ollama. The model and endpoint are fixed in this prototype, so entering a cloud-model URL is not an accidental configuration path. Browser saving is opt-in; the application server does not write incoming notes to disk.
The model does meaningful work here. “Not done: check the spelling” is an unfinished action despite containing “done.” “Cannot finalise the head count until the guests reply” is a dependency without the word “waiting.” A simple keyword sorter is a useful baseline, but it misses those meanings.
The initial garden run put “Still need to check whether the balcony gets four hours of sun” in the waiting category. That was an unfinished action, not an outside dependency. The prompt was clarified, and the original mistake is preserved in the repository rather than replaced by the later result.
That error also explains the dropdowns and “Start here” control. A schema can prevent invented references; it cannot prove that a category is right or that the chosen action suits the person taking the break.
The sample labels and actual outputs are included in evidence/evaluation.json. This is a disclosed development sanity check, not an independent benchmark or a general accuracy claim. The garden example informed the prompt. Categories can still be wrong on unfamiliar or ambiguous notes.
The open-weight model runs on the same device as the notes. Once the runtime and weights are downloaded, this application does not need an inference subscription or a remote model account. The prompt, schema, validation and examples can all be inspected together.
That is useful for a tool handling unfinished thoughts: the default inference path does not require sending those thoughts to a model provider. It also makes the implementation small enough to understand and change. A developer can replace the model deliberately in the source and rerun the examples, rather than depend on an opaque hosted workflow.
There are tradeoffs. The tested model is about 4.7 GB, needs local resources and can start slowly. Saved browser cards are not encrypted. Ollama is a separate application with its own logging and storage behavior. This prototype makes a narrow local workflow inspectable; it does not promise universal privacy or perfect understanding.
The demonstrated payoff is a repeatable return point. Six loose notes become a card, and that card survives clearing the active workspace because the user chose to save it. The app does the sorting; the person only has to review the result and pick the step that feels manageable.
No usability study or measured minutes-saved claim is attached. This is a working prototype, and its next useful test is whether real people find that return card more helpful than reopening their original notes.
No partner prize category is claimed. This entry uses local Qwen2.5 and Ollama, with no paid cloud service.