Skip to content

Open to work · Stuttgart or remote

Gloria Galasso NLP and LLM data scientist

I teach machines to read, listen and (sometimes) explain themselves.

currently obsessed with:text classification, LLM fine-tuning, speech & health, information extraction, interpretability

  • Stuttgart, Germany
Portrait of Gloria Galasso

// about

The short version

NLP and LLM data scientist with an M.Sc. in Computational Linguistics and hands-on Python experience in text classification, information extraction and fine-tuning.

Experienced in building tested, reproducible ML pipelines on AWS, and in turning ambitious research ideas into working tools together with technical and domain experts.

Her path: linguistics → international development → computational linguistics. It explains why she cares as much about the people behind the data as about the F1 score.

  • 7,389

    patient videos in one benchmark

  • 8

    ML approaches compared

  • 5

    languages (one is still loading)

  • 1.3

    M.Sc. thesis grade (1.0 = best)

// training_log.txt

Where she has been deployed

Newest first. From EU policy rooms to multimodal LLMs for clinical research.

  1. Industry

    – (Jun 2026 – Nov 2026)

    Clinical AI/ML Intern

    Modality.ai

    San Francisco, USA (remote)

    Multimodal LLMs for ALS assessment from speech and facial video.

    • Benchmarked 8 ML approaches on 7,389 patient videos; Qwen2.5-Omni embeddings outperformed clinical speech measures (balanced accuracy 0.72 vs. 0.66).
    • Fine-tuned Qwen2.5-Omni-3B and explored LoRA and zero-shot prompting for video captioning.
    • Built a tested Python pipeline (PyTorch, scikit-learn, pytest) on AWS with Git.

    Models that listen and watch faces. Multitasking, finally.

  2. Research

    – (Mar 2026 – Sep 2026)

    Research Assistant

    University of Stuttgart

    Computational Social Science Department, Stuttgart, Germany

    Named-entity extraction from biomedical patents and research papers.

    • Built and evaluated named-entity extraction pipelines with GLiNER on biomedical patent claims and research papers.
    • Ran error analysis on missed entities and boundary errors.

    Spent quality time with patent claims so you do not have to.

  3. Policy & comms

    – (2022 – 2023)

    Project Coordinator

    GOPA Com.

    Brussels, Belgium

    EU and public affairs & communications projects.

    Fluent in "EU policy", the hardest low-resource language.

  4. Policy & comms

    – (2020 – 2021)

    Project Manager

    WONCA Europe / EPR

    Brussels, Belgium

    European healthcare communications & policy support.

    Healthcare policy, long before healthcare data.

// experiments

NLP & speech projects

Two experiments she is happy to talk about for far too long.

2025 – 2026Grade 1.3

Text Classification for Alzheimer’s Disease Detection

M.Sc. thesis · University of Stuttgart

  • Fine-tuned BERT-based transformers to classify spontaneous-speech transcripts, and compared them with logistic regression on linguistic features engineered with the ELFEN library.
  • Interpretability: applied Integrated Gradients to identify which words drove the transformers’ predictions, and compared them with the most predictive engineered features.
  • Evaluated cross-dataset generalization and subgroup robustness.
  • BERT
  • ELFEN
  • Integrated Gradients
  • Interpretability

Peeked inside the black box. It had opinions.

2025

Infant Cry Audio Classification

Speech technology project

  • Extracted MFCC and log-Mel spectrogram features with Librosa and trained 1D CNNs (TensorFlow/Keras) for infant-cry classification.
  • Librosa
  • MFCC
  • 1D CNN
  • TensorFlow/Keras

The loudest dataset, with zero complaints about the labelling.

// model_card.py

Technical skills

The architecture, in four modules. Plus a multilingual head, because why stop at Python.

nlp_and_llms = [ NLP & LLMs

  • LLMs
  • Generative AI
  • Transformers (BERT, Qwen)
  • Text classification
  • Information extraction / NER
  • LoRA fine-tuning
  • Prompt engineering

machine_learning = [ ML

  • Python
  • PyTorch
  • scikit-learn
  • Hugging Face Transformers

speech_audio = [ Speech / Audio

  • Librosa
  • Praat
  • MFCC
  • Mel-spectrograms

tooling = [ Tools

  • AWS
  • Git / GitHub
  • pytest
  • Poetry
  • Linux
  • LaTeX

A multilingual model, naturally

Five supported languages. German is still being fine-tuned.

  • ItalianNative
  • EnglishC1
  • GermanB2 · B2.2 in progress
  • FrenchB1
  • SpanishB1

// pre-training

Education

Three degrees, three countries, one consistent curriculum: language, and what people do with it.

Oct 2023 – Jul 2026

Master of Science in Computational Linguistics

University of Stuttgart

Stuttgart, Germany

1.7 (German scale, 1.0 = best)

  • M.Sc. thesis 1.3; research module 1.0.
  • NLP & ML: Natural Language Processing, Text Technology, Information Retrieval and Text Mining, Computational Semantics, Deep Learning for Speech and Text Processing.
  • Speech & health: Speech Analysis for Health Care Applications (1.3), Current Topics in Speech Technology (1.3), Topics in Phonetics (1.3).

Sep 2019 – Apr 2022

Master of Arts in International Development

University of Kent, Brussels School of International Studies

Brussels, Belgium

Merit

  • Coursework in international human rights law and health care in humanitarian disasters.

2016 – Jul 2019

B.A. in Linguistic Sciences for International Relations

Università Cattolica del Sacro Cuore

Milan, Italy

110/110 cum laude

  • English, German and Spanish, general linguistics.

// inference

Ask the model about Gloria

Not a real LLM, just a tiny (and honest) one. Pick a prompt and watch it stream. Every answer comes straight from her CV.

Choose a prompt:

Waiting for a prompt▍

// deploy_gloria.sh

Ready to ship her to your team?

Looking for an NLP / LLM data scientist who writes tested code, explains her models and speaks five languages? Say hello. She replies faster than a cached response.