Welcome
Welcome to my Website! I’m Miquel, currently driven by a profound aspiration to further my knowledge through the pursuit of a Ph.D., with particular focus on Large Language Models (LLMs) and distributed AI systems. I’m passionate about engaging in cutting-edge research at the intersection of machine learning and distributed systems, pushing the boundaries of technological innovation to shape the future of AI. I’m excited to explore collaboration opportunities and make meaningful contributions to the field. Don’t hesitate to connect with me on LinkedIn or at sirera.m@northeastern.edu if you find our work intriguing, wish to explore partnerships, or if you have any inquiries. Let’s work together to advance the next generation of intelligent systems!
Education
Universitat Politècnica de Catalunya | BSc on Data Science | 2019 - 2023
Northeastern University | PhD on Computer Engineering | 2024 - Current
Experience
Abi Global Health | AI Engineer intern | 2022
Northeastern University | UG Research Assistant | 2023
Current projects
- Building Resilience on Distributed Large Language Models
We are actively building upon the JARVIS project—a distributed framework for Large Language Models (LLMs) that intelligently partitions model layers across edge devices with constrained computational resources. Read the original paper in My Research.
JARVIS introduces enhanced resilience to node failures through peer-to-peer communication and layer redundancy, ensuring reliable performance even in challenging conditions. We evaluated this framework with Google’s Gemma LLM (2B parameters), deploying it on 18 software-defined radios in the NSF Colosseum RF emulator and a 7-node Raspberry Pi testbed.
Currently, we are exploring model adaptations that improve resilience to layer loss, as well as techniques to leverage redundancy in middle layers of LLMs for more efficient recovery, robustness, and optimized distributed execution.
- Extension of Communication-Aware DNN Pruning
In this project, we extend the work on communication-aware deep neural network (DNN) pruning. We focus on training networks to be deployed in a distributed fashion, reducing the need for information exchange while maintaining high accuracy. This approach aims to optimize network performance by balancing computational load and communication efficiency, ensuring robust deployment in distributed environments. We jointly prune and place the neurons of CNNs or MLP-based models to ensure efficient edge deployments under heterogeneous network conditions.
- Knowledge Editing in Large Language Models
We are currently exploring more efficient approaches to knowledge editing in LLMs, aiming to challenge the state of the art in this field. Our focus is on developing methods that enable models to update or correct specific facts and behaviors with minimal retraining and computational overhead, while preserving overall performance and generalization.