PauseDock

• 

  Actual PauseDock interface with invented garden notes; local Qwen inference, original wording retained.

Inspiration

Small volunteer and study projects often stop halfway through. A rough note can be enough to remember the work, but it is easy to lose the next step among things already finished, outside dependencies and background context. PauseDock explores a small way to make returning to a task easier without sending those notes to a cloud AI service. This is a prototype idea, not a measured accessibility benefit.

What it does

Paste 2–20 short notes. A local Qwen2.5 7B model categorizes them as next, done, waiting or context and selects one unfinished note as a place to resume. The app copies the original wording into the return card. Users can correct categories, change the starting note, export Markdown or explicitly save a card in their browser and reopen it after clearing the session.

Primary track: Community Impact. The intended use is task handoff and personal organization in small community or volunteer projects. No community deployment, user research, disability benefit or clinical effect has been demonstrated.

How it was built and AI disclosure

The project was created from an empty directory on 24 October 2026. Python's standard library serves a plain HTML/CSS/JavaScript interface on loopback. Ollama runs the existing open-weight Qwen2.5 7B model locally. The model returns note IDs and categories through a JSON schema; application checks reject invented, duplicate or missing references before showing a card.

OpenAI Codex generated and revised the application code, tests, documentation, submission prose and captioned demonstration with user authorization. These are AI-assisted development outputs; they are not a claim that the account holder personally wrote every line or attended the program's workshops. Qwen supplies the actual runtime classification. The model weights and Ollama runtime are credited third-party dependencies. No API credits or cloud inference were used.

Challenges and observed results

An early real model run treated an unfinished sunlight check as waiting. The prompt was clarified so waiting requires an explicit outside dependency, and the UI retained manual correction. Structural validation can preserve the wording and prevent invented references; it cannot prove semantic correctness.

Fifteen deterministic tests passed. After that prompt refinement the local model matched editorial reference categories for all 16 invented development notes, compared with 13 for a weak keyword baseline. These small examples were part of development, not an independent accuracy benchmark. Warm runs took about 3–5 seconds on the test machine; the first model-loading probe took about 146 seconds.

Technical lessons

Constrained model output is useful for separating classification from text generation. Returning references rather than rewritten notes keeps the original wording auditable. It still requires checking both output structure and the meaning of categories, and providing an easy correction path.

These lessons describe the documented AI-assisted development process. They do not assert personal experiences, skills or learning outcomes not supplied by the participant.

What's next

Test with consenting community volunteers; measure whether returning to a task is easier; improve handling of ambiguous notes; evaluate a smaller local model for devices with less memory. None of those studies or improvements is claimed complete.

Provenance and presentation

The same PauseDock prototype was published as a Hacktoberfest Week 3 entry on 24 October 2026: https://dev.to/yansong_li_b682e6380d7a21/pausedock-let-a-local-model-keep-your-place-while-you-rest-1mkn This submission reuses that openly licensed work and its demonstration; it does not claim exclusive development for STEM2Connect.

The 70-second demo is a captioned sequence of actual app screenshots, edited for length rather than continuous real-time capture. It uses invented garden notes and shows the actual workflow. No private notes, music or third-party imagery.

A prerecorded Demo Day presentation is proposed. The organizer's delivery instructions and deadline must still be checked before treating presentation as complete. No attendance, organizer acknowledgment, certificate or award is claimed.

Built With

• javascript
• python