Open to opportunities

Safa Hamreras

@safahamreras

Message

Postdoctoral researcher at DIPC developing sparse autoencoders that cut LLM parameters by 78% for interpretability.

Spain

Message

At Donostia International Physics Center (DIPC), I developed parameter-efficient sparse autoencoders for LLM interpretability, using 78% fewer parameters and achieving lower reconstruction error than standard SAEs on Gemma 2 9B and Qwen3-8B. I also authored research papers on tensor-based compression of LLMs and CNNs.

At Multiverse Computing, I developed a feature-distillation framework that reduced ResNet-34 parameters by 50% with less than 1% accuracy loss, and co-developed a PyTorch library for neural network compression using quantum-inspired tensor networks. My work also includes fine-tuning language models, researching block-pruning methods, and developing an autoencoder-based anomaly detection solution for Bosch.

Experience

Work history, roles, and key accomplishments

Current

Postdoctoral Researcher in Quantum-Inspired Machine Learning

Current

Donostia International Physics Center (DIPC)

Dec 2023 - Present (2 years 10 months)

Developed parameter-efficient sparse autoencoders for LLM interpretability, achieving 78% fewer parameters and 18-26% lower reconstruction error on Gemma 2 9B and Qwen3-8B. Authored research papers on tensor-based compression and co-developed an autoencoder-based anomaly detection solution for Bosch.

Sparse Autoencoders Model Interpretability Tensor Networks PyTorch Machine Learning Research Anomaly Detection

Senior Scientist in Machine Learning (Part-Time)

Multiverse Computing

Apr 2024 - Apr 2026 (2 years)

Developed a feature-distillation framework for CNNs and LLMs, reducing ResNet-34 parameters by 50% with <1% accuracy loss and achieving 2.35x faster optimization. Co-developed a PyTorch library for neural network compression using quantum-inspired tensor networks and researched block-pruning methods.

PyTorch Tensor Networks Model Compression Knowledge Distillation LLM Fine tuning GPT 2 LLaMA Qwen NLP CNN

Postdoctoral Researcher in Machine Learning and Computer Vision

Universidad de León

Mar 2022 - Jun 2022 (3 months)

Developed image classification models to categorize website screenshots for cybersecurity monitoring. Curated and annotated a dataset of website screenshots and mentored students in machine learning.

Machine Learning Computer Vision Image Classification Dataset Curation Mentorship