QuietBatch 用本機模型精簡每週 CSV 審查

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady.

What I Built

QuietBatch handles a small, repetitive chore: deciding where to start reviewing
a weekly CSV. Give it a representative baseline, the new export, and the numeric
columns that matter. It builds a local Isolation Forest and returns a short
review queue, alongside duplicate rows and invalid numeric values.

The project does not repair files. It preserves the original records and shows
their CSV line numbers so a person can investigate the few rows worth examining.

Demo

Watch the 24-second demo. It replays actual command output from the synthetic fixture. It is a rendered demonstration, not a customer recording. The HTML report source is also available.

The fixture has 200 baseline orders and 50 incoming records. An order with
quantity 99,999 is ranked first by the model; a missing quantity and two identical
rows are separately surfaced. None of these examples is customer data.

Code

QuietBatch source, requirements, fixture generator and tests.

How I Built It

The open-source AI at the center is scikit-learn's Isolation Forest, run locally
on numeric rows. It learns how easily records can be isolated from the baseline.
Unusually isolated incoming rows receive higher review priority. The seed is
fixed so this example can be reproduced. Structural checks stay separate from
the learned ranking.

Five functional tests cover the planted demonstration problems, deterministic
ranking, unchanged inputs, malformed CSV, invalid baseline values and readable
report output. The full suite passed locally. No external inference API was used.

Implementation, tests and this write-up were created with OpenAI Codex assistance.
AI involvement is disclosed here and in the repository README.

Why Open Innovation Matters

An open-source model lets the review happen where the data already lives.
After installing the dependencies, the tool needs neither a hosted inference
service nor an account. Its inputs, seed, feature columns and source code are
available for inspection. Someone can replace the model without moving their
CSV to a different provider.

Limits

A review queue is an attention aid, not a fraud detector or an error guarantee.
Valid unusual orders can rank highly; ordinary-looking mistakes can be missed.
The demo has no measured customer time savings, production accuracy or revenue.
The review budget always selects up to that many valid numeric rows, even when
the input is clean. A person makes the final decision.

Overall category only; no partner category is claimed.