Yousef Qaddura

Quantitative Analytics Specialist — Gen AI in Wealth Investment Management

Python, Async, LangGraph, LLM APIs, Arize AX Evaluations

• Develop and evaluate a LangGraph-based agentic RAG system serving 10,000+ Wells Fargo-affiliated financial advisors.
• Build LLM evaluation frameworks measuring response quality, retrieval relevance, generation relevance, faithfulness, hallucination, toxicity, and adversarial robustness to monitor model performance

Quantitative Analytics Program (QAP) Associate — ML for Fraud Detection

Python, Spark, Synthetic Data Generation, Diffusion Models, Transformers

• Developed and benchmarked generative models for highly imbalanced commercial fraud data (approximately 0.07% fraud rate), evaluating synthetic data quality based on downstream fraud detection performance.
• Improved early-stopping robustness using synthetic fraud data, generating approximately $1M in additional fraud detection value.

QAP Associate — Model Risk Validation (Commercial Stress Testing)

Gen AI, Python, Software Development, Statistical Testing

• Built an end-to-end Python model testing platform supporting multiple model types, integrating LLM APIs, backend services, GUIs, and automated analysis reporting.
• Automated model evaluation across sensitivity, shock, backtesting, stability, and hypothesis analyses, using LLMs to interpret results and generate policy-compliant draft reports, reducing cycle time by ~40%.

QAP Intern — Predictive Marketing Analytics (Credit Cards)

Python, SQL, DataRobot, Machine Learning

• Developed and evaluated XGBoost, LightGBM, neural network, and logistic regression models for credit card mailing campaign response prediction.
• Designed and executed an extensive permutation-importance-based feature selection process, reducing a high-dimensional dataset of 2,000+ credit bureau features to a subset of 11.
• Developed a definition-based feature clustering tool to maintain feature diversity and interpretability.

Estimating the Euclidean distortion of an orbit space

(In collaboration with Ben Blum-Smith, Harm Derksen, Dustin G. Mixon and Brantley Vose)

Invariant Machine Learning, Bilipschitz Invariant Theory, Euclidean Contortion

• Developed general tools for bounding the distortion of metric quotients.
• Applied these tools to various families of quotients by groups of Euclidean isometries.

Group-Invariant Max Filtering

(In collaboration with my PhD advisor, Dustin G. Mixon)

MATLAB, Invariant Machine Learning, Probabilistic Methods

• Used linear algebra and Voronoi analysis to obtain sufficient conditions for the stability and injectivity of max filtering, a recently proposed symmetry-invariant data embedding in machine learning.
• Developed Matlab scripts and utilized linear programming to compute the G-Voronoi characteristic, a newly introduced quantity associated to the representation of a finite group and its max filtering theory.

Stable Weighted Phase Retrieval

Invariant Machine Learning, Differential & Semi-Algebraic Geometry

• Generalized the phase retrieval problem to the weighted setting, highlighting its connections to a nearest neighbor problem in cryogenic electron microscopy (cryo-EM), a leading technique in molecular imaging.
• Used Voronoi analysis, semi-algebraic geometry, and differential geometry to establish sufficient conditions for stability in weighted phase retrieval. This serves as a special case of a broader theory on the local regular stability of max filtering, which the work addresses.

Vector Borne Disease Spread DDE Modeling

(in collaboration with my undergraduate mentor Nsoki Mavinga)

Delay differential equations, MATLAB, Linear Stability Analysis

• Reframed a traditional model predicting spread of infectious, vector-borne diseases.
• Employed linearization methods to prove stability conditions.
• Developed MATLAB interactive applications to simulate numerical solutions of various delay differential equations with real-time parameter tuning.

Food Recommendation System

Python, pandas, Latent Factor Model, Collaborative Team-work

• Completed as part of the Fall 2023 Erdos Institute Data Science Bootcamp.
• Implemented a recommender system that uses the Yelp Dataset to suggest new restaurants to the user by using only their numerical ratings of restaurants.
• Tuned for an optimal mixing hyperparameter in a mixture model of a Latent Factorization Recommender System and a robust averaging baseline model.

Stomper & Wombat's Database Design

SQL, EER Diagrams, Collaborative Team-work

• Put together a requirements document which contained well-written & categorized business rules.
• Formulated a conceptual design using an EER diagram and obtained a relational model at the BCNF level.
• Compiled relevant external views and reports for stakeholders.
• Implemented and tested the database in MySQL server.

Palmer Penguins Regression Analysis

R, R Markdown, ggplot, Regression Analysis, Bootstrap, Cross-Validation

• Conducted an exploratory data analysis on the palmer penguins dataset.
• Carried out a best-subsets regression to obtain a preferred model.
• Performed a cross-validation for assessment and interpreted bootstrap confidence intervals.

Discriminating 7 and 9 with LDA

R, R Markdown, ggplot, Dimensionality Reduction, LDA

• Employed multivariate statistical analysis techniques of principal component analysis and linear discriminant analysis to train a classifier for 7 & 9 digit images.
• Assessed an optimal number of components based on three types of error rates: apparent, test and leave-one-out.