Skip to content
View Yuchi-Wang02's full-sized avatar

Block or report Yuchi-Wang02

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Yuchi-Wang02/README.md

Yuchi Wang

I'm an MS student in Business Analytics and Artificial Intelligence at Johns Hopkins Carey Business School (expected August 2027). I graduated cum laude from Ohio State, with a background in accounting and logistics management.

My projects examine the reliability of AI in business analytics: how data definitions shape results, whether model evaluations are meaningful, and whether business claims are supported by evidence.

Washington, DC · Email · Hugging Face

Selected projects

BizHallu · Checking AI-generated business claims

A correct amount can still support an incorrect claim. In one retail example, an LLM copies the right product and amount but assigns it third place; the evidence places it seventh.

April 2011 example: the same product and GBP 4,173.18 appear at rank 3 in the model answer, but rank 7 in the evidence. This is a curated evidence check.

The workflow connects retail transactions, 100 reproducible business questions, computed reference answers, and local model responses. Interactive cases make the errors inspectable. The evaluation uses provisional, AI-assisted span labels without independent human annotation; the featured case is a curated evidence check.

Explore a case · Methods and results · Code

DelaySentinel · Building and evaluating AI on logistics data

A published Llama fine-tune, a reproducible evaluation workflow, and controlled tests of model behavior. The project connects model development with a practical question: what does a strong score actually establish?

The Llama fine-tune and a depth-2 tree both score 100% on the historical 200-row split. A two-field rule reproduces the source labels.

The audit explains the score through a two-field label rule, then tests how predictions respond to changed inputs—including an irrelevant note that flips all 84 negative predictions, while the antonym Early still returns 1 on every row. What the model actually responds to is the project's open question. The same repository carries a real-order study: classical models trained on checkout-time information and tested on 37,702 later Olist purchases.

Explore the project · Read the case study · Olist study · Model card

Background

Experience includes social intelligence analytics at Ipsos in Shanghai, supply chain analysis at SF Express, and work as a Peer Advisor at Ohio State's Office of International Affairs, supporting international students.

Project stack: Python, pandas, scikit-learn, PyTorch, Hugging Face Transformers, pytest, and GitHub Actions.

These personal projects use AI-assisted implementation and review. Each repository documents its methods, evidence, and limitations.

Pinned Loading

  1. bizhallu bizhallu Public

    Exploring whether AI-generated retail claims are supported by source data, with reproducible questions and interactive cases.

    Python 1

  2. delaysentinel delaysentinel Public

    A published Llama fine-tune scores 100% on its historical split. The audit shows why: a depth-2 tree ties it, a two-field rule reproduces all 1,000 labels, and one unrelated note flips 84 of 84 neg…

    Python 1