Prepared for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady.
AI disclosure: Codex designed and implemented this prototype and drafted this article under Alex C. Christou's direction. Alex reviewed and approved the article and demonstration. The demonstration and evaluation use synthetic examples. This article is disclosed as Fully Autonomous.
What I Built
Quiet Queue turns a small list of postponed tasks into batches: Reply, Read, Arrange and Review. Its purpose is deliberately modest. Someone who has a few minutes for correspondence can see the reply tasks together without first designing a complete productivity system.
The application does not connect to an inbox or calendar. It cannot send a message, book an appointment, remove a task or decide that a duplicate should disappear. Every line remains available for review. An exact repeated task gets a note pointing to the earlier line; both copies stay.
That boundary matters. A model that saves a little sorting effort should not acquire permission to act on someone's messages as a side effect.
Demo
The checked 30-second demonstration shows the actual local browser interface, using held screen captures for readability. Five synthetic tasks enter the sorter. Two become Reply tasks, one becomes Read, one becomes Arrange, and “Water the balcony tomatoes” remains in Review. Nothing is sent or deleted.
Watch the unlisted demonstration. This is a demonstration of the local application, rather than a hosted deployment.
Code
Source and reproduction instructions are available under the MIT licence. The repository contains the Python application, interface and dependency versions. No private correspondence or earlier competition project is included.
How I Built It
The model is classical machine learning, rather than a hosted language-model call. Scikit-learn's TF-IDF vectoriser learns word and two-word features from 24 synthetic task examples. Logistic regression learns three categories. The model is fitted in the local process when the application starts.
An input with no known features goes to Review. The application also abstains when the largest class score is below 0.57. That cutoff is a prototype rule, not a calibrated probability or an assurance of correctness. A confident wrong answer remains possible, especially for unfamiliar wording or mixed instructions.
The browser displays returned text using text nodes. The server binds to the loopback interface, accepts only the application route and does not log task contents or write them to disk. These are limited application behaviours, not a claim about every browser extension or operating-system component.
Why Does Open Innovation Matter?
The complete training vocabulary is visible. There is no hidden hosted prompt, subscription or external inference request. Someone can inspect the examples, replace them with their own non-sensitive examples, change the categories and rerun the same local model.
Using an open-source framework also makes it possible to choose a small, understandable model. A larger generative model would not automatically improve this narrowly defined workflow. The useful question is whether the suggestions are easy to inspect and cheap to disregard.
What Has Actually Been Checked?
Six functional checks passed for task retention, duplicate flags, unfamiliar-input abstention, empty input, input limits and inert handling of HTML-like text. A separate set of six assistant-written synthetic task phrasings received the expected categories. These examples are small and close to the training domain. They do not establish general accuracy or independent validation.
No ordinary working-day usage or measured time-saving result has been verified. The current contribution is a working, limited prototype that makes a repetitive decision visible without silently taking over the next action.
Prize Categories
Overall prize only. No partner technology or paid credit use is claimed.
References: scikit-learn text features, logistic regression.
Top comments (0)