Welcome to My Portfolio
Machine learning researcher working on speech and audio systems for mental health insights.
About Me
I'm a machine learning researcher focused on speech and audio for mental health assessment. My recent work improves depression detection with compact neural architectures and multimodal evaluation. I enjoy translating open questions into deployable systems, and I bring experience across Python, PyTorch, big-data tooling, and rigorous model validation. I'm eager to contribute to lab efforts that combine methodological depth with applied impact.
Skills
Mathematics & Statistics
Calculus (single/multivariable), linear algebra, probability and statistical inference, Bayesian modeling, time-series analysis
Programming Languages
Python, R, JavaScript, Java, C++, ABAP
Frameworks & Libraries
PyTorch, scikit-learn, TensorFlow, Transformers, NumPy, Pandas, p5.js
Web & Big Data Tools
HTML, CSS, PHP, Docker, Spark, SQL, AWS EC2, Tableau, Power BI, PowerApps
Fundamentals
CI/CD pipelines, Linux CLI, time-series modeling, model evaluation, data cleaning, stakeholder collaboration
Languages
English (Fluent), German (Native), French (Basic)
Certifications
Professional Scrum Master (PSM-1)
iOS App Development Certificate
Publications
Soft Clustering Anchors for Self-Supervised Speech Representation Learning in Joint Embedding Prediction Architectures Under Review – ICML 2026
Georgios Ioannides, Adrian Kieback*, Judah Goldfeder*, Yann LeCun, Aman Chadha, Aaron Elkins, Linsey Pang, Ravid Shwartz-Ziv
* Equal contribution
Under review at ICML 2026; preprint available on arXiv (January 2026).
DAWZY: AI for Music (NeurIPS 2025 Workshop)
Aaron C Elkins, Sanchit Singh, Adrian Kieback, Sawyer Blankenship, Uyiosa Philip Amadasun, Aman Chadha
Accepted to the NeurIPS 2025 AI for Music Workshop; presenting December 7, 2025.
Research Experience
Machine Learning Researcher I 09/2023 – Present
James Silberrad Brown Center for Artificial Intelligence, San Diego, USA
- Automated and enhanced audio voice processing by implementing advanced neural network models, improving depression detection accuracy.
- Increased macro F1 score by 21% compared to state-of-the-art systems.
- Utilized a server with two A100 to optimize models to be 72% more parameter-efficient than industry standards.
- Optimized and fine-tuned stable diffusion models (LAION 5B dataset), improving performance and efficiency.
- Published research findings on arxiv.org/abs/2409.00391.
- Developed GMM-Anchored JEPA, a self-supervised speech representation learning architecture that fits a Gaussian Mixture Model once on log-mel spectrograms and uses frozen soft posteriors as auxiliary targets, eliminating the need for iterative re-clustering required by HuBERT and WavLM.
- Trained and evaluated on ~50k hours of speech using a DGX B200 cluster (8 GPUs, 1.4 TB VRAM); achieved 28.68% WER on ASR (vs. 33.22% WavLM-style baseline), 67.76% accuracy on emotion recognition, and 64.7% F1 on slot filling.
- Author on resulting paper submitted to ICML 2026 and published as arXiv preprint arXiv:2602.09040.
Data Analyst Research / Graduate Assistant 08/2023 – 05/2024
San Diego State University, San Diego, USA
- Proctored and graded exams for two professors' classes, ensuring timely feedback.
- Automated a web scraping process to collect data using Python.
- Utilized Python and R to enhance data collection and processing efficiency.
Currently Working On
Statistical Modeling of Environmental and Geographic Risk Factors for Lung Cancer
Advisor: Aaron Elkins | In collaboration with SHARP
- Building multivariate regression, generalized additive, and mixed-effects models to identify environmental and geographic factors beyond smoking that are associated with lung cancer incidence.
- Developing geospatial workflows for detecting regional clustering, exposure patterns, and spatial autocorrelation, with careful attention to confounding and missingness.
Speech-Based Prediction of Depressive Episodes in Bipolar Youth
Advisor: Aaron Elkins | Collaboration with Baylor College of Medicine
- Designing an end-to-end pipeline for longitudinal interview speech, including SNV-normalized Mel spectrograms and sequence modeling.
- Planning evaluation of LSTM, Transformer, and JEPA-inspired encoders for early detection of mood-state transitions, with interpretability analyses to support clinical relevance.
Work Experience
Data Science Intern 02/2024 – 05/2024
SANDAG, San Diego, USA
- Visualized geospatial data for San Diego County to track UN Sustainable Development Goals.
- Conducted extensive data cleaning to ensure accuracy of published statistics.
- Automated predictive analytics workflows for toll operations on 90,000+ customer accounts.
Business Scientist Intern / Bachelor Thesis Author 02/2022 – 08/2022
Carl Zeiss AG, Aalen, Germany
- Developed BI reports for 1,200+ machines globally, integrating live data from multiple sources.
- Performed stakeholder interviews and built a forecasting dashboard, improving data-driven decisions.
- Influenced software choice, achieving annual savings of \$30,000.
Projects
Baseball Prediction
- Developed a Dockerized Python tool for statistical analysis and predictive modeling, generating HTML reports.
- Engineered 20+ time series features (SQL, PySpark, MariaDB) and implemented an Elo-ranking system.
- Evaluated six ML models using ROC curves and cross-validation, selecting Logistic Regression for best performance.
Interactive Web-Based Visualization of Public Health Data
- Created interactive web maps with p5.js and Leaflet.js to show surgical site infection data across California.
- Processed 26,000+ rows of public health data, enabling dynamic, user-friendly visualization.
Education
San Diego State University | MS Big Data Analytics
GPA: 3.9
- Master Thesis: “Enhancing Mental Health Diagnosis with Representation Learning: A Transformer-Based Approach”
Albstadt-Sigmaringen University | BS Business Informatics
GPA: 3.5
- Bachelor Thesis: “Conceptual design of a central tool for the globally standardized presentation and analysis of process and machine data”