Maria Magdalena Balos

Summary

Data scientist with a strong focus on applied machine learning, NLP, and generative AI, complemented by a background in UX research. Experienced in training and deploying speech models in production, building RAG systems, and implementing deep learning architectures from scratch. Combines deep ML expertise with a UX research background, enabling technical depth and user-centred thinking, with a pragmatic, committed approach to problem solving, delivering results independently and communicating findings clearly across technical and non-technical audiences.

Experience

Data Scientist — Vocality.ai

• Trained and deployed text-to-speech (TTS) models across multiple languages (en-GB, es-ES, eu-ES, sl-SI), serving end users and enterprise clients including Mercedes-Benz, Heineken, and ONCE, contributing to €10k+ monthly recurring revenue.
• Reduced critical mispronunciation rate in es-ES from ~11% to <0.5% through improved data quality, fine-tuning, and model architecture improvements.
• Automated manual workflows, reducing operational overhead by ~50%.
• Maintained and extended MLOps pipelines on GCP, integrating TTS models into production deployment workflows.
• Built an audio post-processing pipeline assembling multi-segment TTS outputs into complete audio tracks based on timestamps, removing silences, and formatting audio for ads and e-learning applications.
• Generated and validated subtitle outputs from TTS pipelines.
• Built business analytics dashboards in Streamlit to visualise app usage statistics and user spending.

Research Assistant — Museum of Archaeology and Anthropology

• Conducted contextual observation studies, systematically tracking visitor movement, timing, and interaction patterns across exhibits.
• Collaborated in the design of a 32-question structured survey to collect quantitative visitor behaviour data.

UX Researcher — Singer Instruments

• Automated interview transcription and speaker diarization using Whisper and Pyannote, reducing transcription time by ~80% and enabling automatic labelling of user and interviewer speech turns for downstream analysis.
• Designed and led usability studies, user interviews, and A/B testing in laboratory environments with scientists.
• Analysed qualitative and quantitative research data to inform product decisions across a multi-disciplinary R&D team.
• Awarded Intern of the Year 2022.

Projects

Manifold HyperConnections for Computer Vision In progress
arXiv:2512.24880

• Replicating results from the Manifold HyperConnections (mHC) paper and conducting experiments on how this novel architecture can be applied to computer vision tasks.

CON(e)VOLUTION – From LeNet to Vision Transformers
GitHub · Article

• Implemented 10 CNN architectures from scratch in PyTorch — including custom variations of existing architectures — benchmarked and tested on the NEU Surface Defects dataset, with a shared modular trainer and CLI interface.

RAG-Driven Educational Assistant · Master's Dissertation
GitHub

• Ingested ~364h of video content (91 lessons) via Whisper transcriptions into a vector knowledge base.
• Built a full-stack RAG application using LangChain, ChromaDB, and Large Language Models (LLMs) including GPT-4o-mini and Claude-3-Haiku.
• Benchmarked RAG-based vs. non-RAG-based approaches to evaluate and quantify retrieval benefit.
• Implemented gamification in the Flask interface, repurposing the benchmark as a student leaderboard to compare human performance against the RAG/LLM system.

Ryanair Timecapsule
GitHub

• Reverse-engineered the Ryanair API to enable systematic daily collection of flight price data across two endpoints. Released as an open-source Python package with tests and CLI scripts to facilitate data collection for researchers interested in price prediction and forecasting.

K-Means for Colour Palette Generation
Article

• Applied K-Means clustering to extract and visualise dominant colour palettes from images, deployed as a Flask web application for interactive data visualization.

Education

Master's in Deep Learning & Generative AI · Datamecum, Madrid

• Dissertation: RAG-Driven Educational Assistant — Q&A and knowledge testing system for DS students.
• Coursework: computer science, neural networks, transformer architectures, generative models, NLP, computer vision.

Intensive Program in Data Science · Datamecum, Madrid

• Won 1st place in the 3rd Datamecum Datathon (Kaggle-style competition) — conducted exploratory data analysis and built a binary classification model (AUC 0.9851) using an ensemble of Random Forest and XGBoost.
• Coursework: statistical analysis, modelling techniques, machine learning algorithms, EDA, supervised and unsupervised learning, outlier detection.

Master's in Interaction Design & UX · Universitat Oberta de Catalunya

Bachelor's in Graphic Design & Digital Creations · Universitat Oberta de Catalunya

Skills

Machine Learning & Artificial Intelligence (ML & AI): Deep Learning, Natural Language Processing (NLP), Speech Models (TTS, ASR), RAG, LLMs, Computer Vision, CNNs, Transformers, Generative AI, Decision Trees, Gradient Boosting (XGBoost), Linear Regression

Frameworks & Tools: PyTorch, Hugging Face, LangChain, ChromaDB, Scikit-learn, Pandas, NumPy, Streamlit, Flask

Cloud & Infra: Google Cloud (GCP), MLOps, Version Control Systems (Git), Docker

Programming Skills: Python, SQL, Bash

Other: Data Analytics, Data Visualisation, Speaker Diarization, Whisper, UX Research, Agile, Teamwork, Unit Testing

Certifications & Continuous Learning

• Coursera (Andrew Ng): Advanced Learning Algorithms, Supervised Machine Learning
• fast.ai: Practical Deep Learning
• Hugging Face: NLP with Transformers, Audio Transformers
• 300+ LeetCode problems solved · Advent of Code 2024
• Active member, Cambridge Programmers Study Group (weekly meetup)

Languages

English (full professional) · Spanish (native) · Romanian (native) · French (basic)