Guus Bouwens

Programming

• Front-End: HTML, CSS, JavaScript

• Back-End: Python (NumPy, Pandas, Scikit-Learn, TensorFlow, Keras), R, PHP, C#, Java, C++

• Database: Relational database structures, SQL, Dbeaver, Azure functions

• Cloud Platforms: AWS, Microsoft Azure, GCP

• Analysis: PowerBI, Tableau, Excel

• Version Control: Git, Jenkins, Docker, CI/CD

Computer Science

• Core Concepts: Database Fundamentals, Data Structures & Algorithms (DSA)

• Machine Learning: Supervised & Unsupervised Learning, Reinforcement Learning

• Deep Learning: Neural Networks, Natural Language Processing (NLP), Transformer Models, LLMs (e.g., Llama, GPT, Claude)

• Artificial Intelligence: Sentiment Analysis, AI API's (Hugging Face, OpenAI), Automation, Fine-Tuning LLMs, RAG

Economics

• Macro & Microeconomics: Economic Theory, Policy, and Market Structures

• Quantitative Finance: Quantitative Methods, Volatility Evaluation

• Ethics: Ethics in AI, Big Data, Finance & Economics

• Investment & Portfolio Management: Investment Strategies, Risk Assessment

Mathematics

• Core Concepts: Calculus (Analysis, Derivatives, Integrals), Linear Algebra, Multivariate Statistics

• Probability & Statistics: Econometrics, Statistical Modeling, Time Series, Fourier Analysis

• Numerical Methods: Numerical Simulations, Optimizations, Bootstrapping

Communication

• Academic: Academic Reporting, Presenting, LaTeX

• Interpersonal: Interpersonal Skills, Collaboration, Teamwork

• Management: Project Management (Agile/Scrum, Jira), Leadership, Communication

• Tools: Slack, Microsoft Teams, Microsoft Powerpoint/Word/Spreadsheets

Language

• Fluent: English and Dutch

• Beginner: French, German, and Spanish

AI Engineering Intern @ McDermott International, Ltd

October 2024 untill April 2025 in The Hague

• Optimized RAG System: created an evaluation dataset, implemented custom evaluation metrics, tested and implemented the best chunking methods, and added metadata for improved retrieval and response quality. Integrated hybrid retrieval and reranking techniques. All of this was done in a secure environment, ensuring no data leakage.
• Worked with Git, Python, Viktor CLI and Azure for CI/CD. Followed the Scrum Agile working method.

AI Engineering Intern @ Invest-NL

April 2024 untill September 2024 in Amsterdam

• Fine-tuned and integrated Large Language Models (LLMs) like Llama-3.1 for company classification, optimizing processes for strategic decision-making.
• Collaborated with teams in the energy transition and sustainable investment sectors.
• Utilized tools like Git, Python, SQL, Power BI, and Azure while following Agile Scrum and using Microsoft Teams.

Data Science Intern @ Talpa eCommerce

June 2023 untill September 2023 in Amsterdam

• Enhanced auction website recommendation systems by implementing advanced similarity measures and a custom Keras transformer.
• Conducted Fourier transformations on big data to generate actionable customer insights.
• Worked with Python, SQL, Jenkins, Docker, and AWS in a CI/CD environment while following Agile Jira and using Slack.

MSc. Data Science @ Texas Tech University

May 2025 until April 2026

• Accepted a scholarship and will be attending on campus.

BSc. Econometrics and Data Science @ VU Amsterdam

September 2020 until April 2025

GPA: 7.5/10 or 3.5/4

• Minor in Big Data (Scored 8/10)
• Led several capstone projects with best-in-class grades.
• Courses: multivariate statistics, linear algebra, macro economics, financial econometrics, machine learning, data analysis, and database management.

BSc. Applied Physics @ TU Delft

September 2019 until July 2020

Withdrew to follow my passion with data-driven causality and programming.

Pre-University Education @ Lyceum Sancta Maria Haarlem

September 2012 until July 2018

• Majored in Physics, Chemistry, Biology, Economics, Computer Science, and the highest level of Mathematics.
• Finished with a capstone imitating quantum tunneling, scoring a 8.5/10.

AI

• Mathematics for Machine Learning @ Imperial College London

• Machine Learning @ University of Washington

• Deep Learning @ DeepLearning.AI

• Natural Language Processing @ DeepLearning.AI

• Artificial Intelligence in Healthcare @ Stanford University

• Ethics of AI @ Politecnico di Milano

Data Science

• Azure Data Scientist Associate @ Microsoft

• Applied Data Science with Python @ University of Michigan

• Data Science using R @ Johns Hopkins University

• SQL for Data Science @ University of California, Davis

• Data Structures and Algorithms @ University of California, San Diego

• Recommender Systems @ University of Minnesota

Analysis

• Excel for Data Analytics and Visualization @ Macquarie University

• Data Analytics @ Google

• Data Analysis using PySpark @ Coursera

• Data Visualization with Tableau @ Duke University

• Data-Driven Decisions with Power BI @ Coursera

Finance

• Investment Management with Machine Learning @ EDHEC Business School

• Fintech @ University of Pennsylvania

• Finance & Quantitative Modeling @ UPenn Wharton

• Financial Engineering and Risk Management @ Columbia University

• Financial Markets @ Yale University

Other

• Project Management @ Alphabet

• Aeronautical Engineering @ TU Delft

University (grade)

Show

• Probability and Inference Game Using the Newsvendor Problem (10/10): I applied theoretical concepts from the Newsvendor Problem (NVP) to optimize order quantities for a European bakery chain. Using real-world data, I modeled profit functions while incorporating holding costs, and derived the optimal order quantities using parametric and non-parametric methods. Through Monte Carlo simulations, I compared the performance of these methods and analyzed the strengths and weaknesses of each. The results indicated that parametric methods outperformed non-parametric ones when the underlying demand distribution was correctly specified. Additionally, I performed a detailed time series analysis on bakery demand data to identify trends and patterns, especially focusing on the impact of weekends on demand variability.

• Classifying Financial News Articles Using Machine Learning (9.5/10): I worked on classifying financial news articles from the Reuters-21578 dataset, specifically identifying articles related to crude oil. The project involved pre-processing textual data through tokenization, stemming, and the removal of stop words. Using a Term Frequency-Inverse Document Frequency (TF-IDF) matrix and Latent Semantic Analysis (LSA) for dimensionality reduction, I built and compared several machine learning classifiers, including Logistic Regression, k-Nearest Neighbors (k-NN), and Naive Bayes (NB). I also implemented the AdaBoost algorithm to boost the performance of NB classifiers. The results showed that Logistic Regression achieved the highest recall, while k-NN outperformed other models in precision. This project provided valuable insights into document classification techniques in natural language processing and their application to financial data.

• Factor Analysis of Musical Attributes Across 50 Artists (8.5/10): This project involved constructing an orthogonal factor model (OFM) to uncover underlying patterns within a dataset of musical attributes, including energy, danceability, and popularity. By applying a VARIMAX rotation, I identified three key factors—calmness, non-danceability, and studio production quality. These factors provided insights into genre distinctions, artist positioning, and the relative influence of different musical characteristics. The analysis was further visualized by plotting artists on an energy-danceability spectrum, offering valuable information for music industry stakeholders and recommendation systems.

• Stochastic Volatility and CAPM with Time-Varying Beta (8.5/10): This project involved modeling volatility dynamics for the S&P 500 using a Stochastic Volatility (SV) model and comparing it to traditional GARCH models. The SV model demonstrated high persistence and volatility clustering in financial time series, providing valuable insights for risk management. The project also incorporated a dynamic version of the Capital Asset Pricing Model (CAPM) with time-varying beta to capture evolving market risks for Microsoft, Bank of America, and Exxon Mobil. By employing Maximum Likelihood Estimation (MLE), I assessed changes in market exposure over time and explored the benefits of dynamic risk modeling for portfolio management.

• Ethics of Supplying Drugs for Lethal Injections: A Moral Dilemma (8.5/10): In this essay, I explored the ethical considerations surrounding pharmaceutical companies supplying drugs for lethal injections used in executions in the United States. The analysis drew from various ethical frameworks, including utilitarianism, Kantian ethics, and Aristotelian perspectives, to examine whether it is morally justifiable for pharmaceutical companies to provide drugs for executions. The essay highlighted the conflict between the human right to life and the government's effort to make executions as humane as possible. The ethical dilemma was further complicated by the potential consequences of refusing to supply drugs, which could lead to less humane alternative execution methods. Ultimately, the essay provided a nuanced evaluation of this complex ethical issue from multiple moral viewpoints.

• Time Series Analysis (10/10) and Vector Error Correction Models (8/10): In these two projects, I conducted a detailed time series analysis of macroeconomic indicators, focusing on French and Euro area data. The first project involved modeling GDP growth, inflation, and interest rates for France using vector autoregressive (VAR) models, performing hypothesis tests, and assessing model stability. The second project expanded on this work by analyzing Euro area time series data, incorporating housing wealth and exploring cointegration relationships using Vector Error Correction Models (VECMs). I applied advanced statistical methods to uncover long-term equilibrium dynamics and short-term adjustment processes.

• Volatility and Risk Forecasting Using GARCH and GAS Models (8/10): I developed models to predict the volatility and risk measures of daily returns for financial assets, including a stock (ASML), a cryptocurrency (Bitcoin), and an exchange-traded fund (S&P 500 ETF). Using both GARCH and Generalized Autoregressive Score (GAS) models, I generated 1-day-ahead forecasts for volatility and conducted backtesting to assess the accuracy of these models in predicting Value at Risk (VaR) and Expected Shortfall (ES). The analysis revealed that GARCH models performed well for low-volatility data like the S&P 500, while GAS and HARCH models better captured the volatility dynamics of highly volatile assets like Bitcoin. This project provided insights into the relative strengths of GARCH and GAS models for financial risk management across different asset classes.

• Predicting Law Graduate Salaries Using OLS and k-NN Regression (8/10): I developed and compared regression models to predict the starting salaries of law graduates based on factors such as law school rank, LSAT scores, GPA, and the cost of education. Using Ordinary Least Squares (OLS) and k-Nearest Neighbor (k-NN) regression, I analyzed the relationship between law school prestige and graduate salaries. The project also incorporated Monte Carlo simulations to assess uncertainties in the results. Ultimately, the KNN model outperformed OLS in several instances, demonstrating better predictive accuracy for this dataset. This analysis provided valuable insights into the factors influencing graduate earnings and the benefits of non-parametric methods like KNN in regression analysis.

• Bootstrap Methods for Time Series Analysis (8/10): I explored the performance of various bootstrap methods in time series analysis, particularly focusing on heteroscedasticity in data. Through an in-depth comparison of methods such as the residual bootstrap, wild bootstrap (with both normal and Rademacher distributions), and pairs bootstrap, I evaluated their effectiveness in hypothesis testing, particularly under heteroscedastic conditions. Results indicated that the wild bootstrap with Rademacher distribution consistently provided the most accurate results. This project offered valuable insights into the reliability of bootstrap methods in econometric models, particularly when dealing with time series data.

• Modeling Volatility in the German Stock Market (8/10): I analyzed the volatility dynamics of the German stock market, focusing on the DAX index using GARCH and GJR-GARCH models. The goal was to identify patterns of volatility clustering and assess risk using various Value-at-Risk (VaR) measures. By comparing multiple GARCH models (GARCH(1,1), GARCH(2,1), and GARCH(2,2)), I identified GARCH(1,1) as the most suitable model. The project further involved portfolio management using multivariate GARCH models, where I estimated optimal portfolio weights for DAX and HSI stocks under different market conditions. Additionally, I explored time-varying conditional correlations using the CCC model and evaluated how model choice impacts risk assessment and investment strategies.

Professional

@ McDermott International, Ltd

• RAG optimization: implemented custom evaluation metrics, tested and implemented the best chunking methods, and added metadata for improved retrieval and response quality. Integrated hybrid retrieval and reranking techniques. All of this was done in a secure environment, ensuring no data leakage.

@ Invest-NL

• Evaluated and Fine-tuned LLMs: I tackled the challenge of optimizing company classification using state-of-the-art Large Language Models (LLMs). This project focused on evaluating and fine-tuning models like GPT-4o, Claude 3.5 Sonnet, Llama-3.1, and Mistral Large 2 to categorize companies in innovative sectors such as Deep Tech, Life Sciences & Health, and Agrifood. The Llama-3.1–70B-Instruct model demonstrated the highest accuracy and efficiency for the task, surpassing traditional machine learning methods. Additionally, I fine-tuned a smaller Llama-3.1–8B model using Direct Preference Optimization (DPO), achieving a boost in speed and classification accuracy. This work highlights the potential of fine-tuned models for enhancing AI-driven investment decisions. I also wrote an article about this project!

• Developed Dashboards for Company Insights by Focus Area: I created comprehensive dashboards that provided teams with a clear overview of companies categorized by sector and subsector. These dashboards included detailed company descriptions and vital financial information such as growth stage, investor details, funding history, and the number of employees. This centralized system offered teams an efficient tool for tracking companies across Invest-NL’s distinct focus areas, enabling better decision-making and resource allocation.

@ Talpa eCommerce

• Built a Keras Transformer-Based Recommendation System: I developed a dynamic, transformer-based recommendation engine using Keras. This system leveraged customer behavior data to provide highly personalized product recommendations. By incorporating deep learning techniques, the model continuously adapted to new data, improving the accuracy and relevance of the recommendations over time.

• Improved Product Matching Using Similarity Measures: I focused on enhancing product matching through advanced similarity measures. I utilized TensorFlow with web scraping to improve image-based product similarity, applying cosine similarity metrics for targeted marketing strategies, including product-to-product and geospatial recommendations. The implementation of these similarity models led to more accurate product matches, driving improved customer engagement, higher recommendation relevance and thus higher margins.

• Analyzed Customer Seasonality with Fourier Transformations: I utilized Fourier transformations to analyze large-scale customer bidding data and uncover seasonal patterns in customer behavior. By examining over two years of auction data, I identified the intervals at which customers returned to bid on similar products. This method allowed me to compute a "Trendiness Score" for each product, combining returning customer percentages, time-weighted intervals between purchases, and dominant purchase frequencies. These insights provided a data-driven approach to predict customer re-engagement and optimize product recommendations based on seasonality trends. For example, the analysis showed that, on average, customers tend to return approximately every 30 days to bid on essential items like toilet paper, making it possible to adjust the timing of product recommendations to match their purchasing cycle.

Independent

• Crypto Dashboard: A real-time cryptocurrency dashboard showing price and social media sentiment analysis. It's main addition to other sites is the constant analysis of new youtube videos.

• Arcade Space Shooter: A dynamic space shooter game built with vanilla JavaScript, HTML5, and CSS3, featuring progressive difficulty scaling and engaging power-up mechanics. Players navigate a spaceship through an asteroid field while managing weapon heat levels and collecting power-ups like triple shots, shields, and unlimited fire modes. The game implements a unique scoring system where both difficulty and score multipliers increase every 30 seconds, encouraging longer survival times and higher scores. Technical features include smooth animations, collision detection, local storage for high scores, and responsive controls - all built without external libraries. The project demonstrates proficiency in game state management, object-oriented programming, and browser performance optimization while delivering an engaging user experience..

• Open-World Automation Survival Game: I am currently also developing a bigger game from scratch using C# and Unity, set in a post-civilization world. The game blends elements from Factorio, Minecraft, and Valheim, featuring survival, combat and automation mechanics. My goal is to create an immersive experience where players build, manage resources, and interact with dynamic in-game economies, all while navigating the challenges of a world rebuilding from collapse.

• I built this website from the ground up using PHP, JavaScript, HTML, and CSS. The site is fully custom-coded, featuring dynamic functionality and a clean, navigable, and responsive design.

• Harvard ML Capstone: I developed a movie recommendation system using the MovieLens 10M dataset, focusing on predicting user ratings for movies. After performing data preparation and analysis, I built three models: a naive model predicting the same rating for all movies (RMSE: 1.06), a movie effect model incorporating movie-specific averages (RMSE: 0.94), and a movie + user effect model that achieved the best performance with an RMSE of 0.87.

• I have also competed in some kaggle competitions and am active on leetcode.

• More projects and code can be found on my github page.

• Fine-Tuning LLMs for Investment-Grade Company Classification: A Frontier Model Comparison