Crystal(Tianhui) Hu

Grew up in Australia & Aussie citizen with Japan's Permanent Residency, results-driven AI and data scientist with over six years of experience specializing in credit risk, fraud detection, and RAG in natural language processing. Global perspective enriched by five professional years in Sydney with local English and two years in Tokyo, collaborating directly with Japanese stakeholders in their native language. Expertise includes building predictive models and engineering advanced NLP solutions, highlighted by the development of a patented RAG-based 'Evidence Tracker' that effectively eliminates LLM hallucinations to ensure reliable information delivery. Thrives in cross-functional environments with a strong commitment to aligning technical solutions with business objectives, leveraging a comprehensive, user-focused mindset to shape product direction and enhance tool robustness through innovative NLP methods.

Kanji Generation with Stable Diffusion (NLP Research Project, Aug 2025 - Present)
Goal: The goal o f this project was to train a stable diffusion model t o generate novel Japanese Kanji characters from English definitions, reproducing the viral experiment that demonstrated Al's ability t o "hallucinate" new cultural symbols for modern concepts like "YouTube", "Gundam", and "Elon Musk" that don't have existing Kanji representations.

• Data engineering and dataset creation: built a comprehensive dataset o f 6,410 Kanji characters by parsing KANJIDIC2 XML files t o extract English meanings and KanjiVG SVG files for stroke data, converted vector SVG drawings to 128x128 pixel images with pure black strokes (#000000) o n white backgrounds, ensuring no stroke order numbers were rendered, created complete metadata mapping with 16,692 English meanings (average
  2.6 per Kanji) and proper text prompts for training, implemented a quality control pipeline t o validate image consistency and remove artifacts
• Generation and results: Developed an advanced concept generation system capable of creating Kanji for modern concents like "¡Phone " "Bitcoin " "Netflix " "Tesla. " "Instagram " "COVID-19." and "artificial intelligence"
  

Loan Prediction (Commonwealth Bank, Feb. 2022 – Apr. 2022)

Goal: The goal of the loan prediction model was to accurately predict whether a loan would be successfully approved or not based on the selected metrics. By utilizing feature selection techniques and building both Random Forest and logistic regression models, we aimed to achieve high prediction accuracy and validate the performance of the models.

• For feature selection, selected the most meaningful metrics according to business scope and end goal: LoanId, gender, marriage status, dependents, education, applicantIncome, coapplicantIncome, LoanAmount, LoanAmountTerm, credithistory.
• Filled missing values using mode or mean value, and split the data into 70% training data and 30% testing data.
• Built a Random Forest Model and fit it to the dataset, achieving an accuracy of 77.2% which is fast and simple enough to predict if a loan would be successfully approved or not.
• Built a logistic regression model and fit it to the dataset, achieving an accuracy of 82.2% to compare the prediction accuracy and validation score.

Probability of default (PD) model development (Commonwealth Bank, Apr 2023 - Sep 2023)
Goal: developed a probability o f default (PD) model t o estimate the likelihood o f default for credit card holders based on historical data and relevant features
• Developed a probability o f default (PD) model t o estimate the likelihood o f default for credit card holders based on historical data and relevant features, such as customer demographics, credit history, and financial Indicators
• Conducted exploratory data analysis to understand the underlying patterns and relationships i n the dataset
• Preprocessed and cleaned the data by handling missing values, outlier detection, and feature engineering
• Selected appropriate machine learning algorithms, including logistic regression, decision trees, and gradient boosting, to build and train the PD model
• Evaluated the model's performance using various metrics, such as accuracy, precision, recall, and F 1 score and fine-tuned the model t o optimize its predictive power
• Incorporated model interpretability techniques, such as feature importance analysis and partial dependence plots, to gain insights into the factors driving the default probability
• Collaborated with stakeholders, including risk management teams, and business analysts, t o validate and refine the PD model's performance, and ensure its alignment with business objectives