SaleSense

A tool that scores resale listings and generates clear, higher-converting product descriptions.

SaleSense helps sellers to improve second-hand products by scoring the description and generating a new better description. It’s built to be fast, simple, and practical for real resale workflows.

Problem

Most resale platforms depend heavily on listing quality, but many users don’t know what details to include or how to write descriptions that sell. SaleSense makes that improvement automatic: score → feedback → improved description.

Dataset imbalance

The dataset was extremely imbalanced (sold vs. not sold), which made the classifier approach unreliable.

Distribution of “sold” status and the initial baseline attempt.

Making the dataset workable

To train and validate models more reliably, we down sampled the negative class to create a more balanced working set.

Downsampling strategy used to reduce imbalance and stabilize training/evaluation.

LLM feedback loop

For rewriting and feedback, we use an LLM-driven feedback loop to produce clearer and more complete descriptions. A LightGBM classifier scores the inputed description and provides a target signal, guiding the LLM to iteratively rewrite the text to maximize the predicted “sellability” score.

LLM feedback loop used to produce improvements and guidance.

Design + build

The interface was designed to be minimal and fast: paste a description, generate feedback, and get an improved version immediately. The project is implemented as a lightweight web app (Streamlit) so it’s easy to run and demo.

Repository

Code: https://github.com/PedroTajia/AISC_SaleSense

References