Pascual Merita

I’m currently a Co-Founder at Arthos, building Mozart AI. Mozart AI is a Digital Audio Workstation (DAW) with built-in AI Co-Production. We empower what we call the ‘Hybrid Musician’ — an artist driven by human creativity and assisted by AI. Human work will slowly shift toward high-level tasks while AI handles (more) routine operations. At Arthos, we want artists to stay in their creative flow, while AI takes care of the tedious components and sometimes even throws in a spark of inspiration. We are starting with music.

Previously I was an AI Research Engineer at Huawei, where I focused on Retrieval Augmented Generation (RAG) under the supervision of Pavlos Vougiouklis and Jeff Pan. If you’re interested in learning a bit more about me, please take a look at some of my blog posts.

Prior to joining Huawei, I earned my Master’s degree in AI from The University of Edinburgh. My research there focused on improving the efficiency of Vision Transformers through token pooling techniques, achieving up to 1.4x faster inference times while maintaining comparable accuracy. This work was conducted under the guidance of Edoardo Ponti and Piotr Nawrot. During my time in Edinburgh, I also co-founded and was president of Edinburgh AI, the university’s AI society. With now over 400 members, talks by the likes of DeepMind, and a yearly conference where students can present their work, it has grown into a thriving community that I am proud to have started.

Before that, I earned a BSc in Computer Science from King’s College London. My thesis focused on ‘Similarity-based Music Recommendation Systems’, where I employed various CNNs to recommend songs similar to a given input. My official supervisor was Jeroen Keppens, while I also received valuable guidance from my unofficial advisors, Timothy Greer (Amazon Music) and Minz Won (Suno AI).

Overall, I would say I am driven by two different kinds of projects: One kind consists of projects that have a significant, tangible impact on other people’s lives. The other kind are those which pose questions that motivate me from a philosophical perspective. I am happiest when I can work on a project that ticks both boxes. Learn more about what I am optimising for in life right now in this blogpost.

def get_my_email(self):
    email = "definitely.not.a.bot@gmail.com"
    if self.not_a_bot:
        name = "Pascual Merita Torres"
        parts = name.lower().split()
        email = f"{parts[0][0]}{parts[1]}{parts[2]}@gmail.com"
    return email

selected publications

arXiv

GeAR: Graph-enhanced Agent for Retrieval-augmented Generation

Z. Shen^*, C. Diao^*, P. Vougiouklis^*, P. Merita^*, and 7 more authors

Dec 2024

Abs HTML PDF

Retrieval-augmented generation systems rely on effective document retrieval capabilities. By design, conventional sparse or dense retrievers face challenges in multi-hop retrieval scenarios. In this paper, we present GeAR, which advances RAG performance through two key innovations: (i) graph expansion, which enhances any conventional base retriever, such as BM25, and (ii) an agent framework that incorporates graph expansion. Our evaluation demonstrates GeAR’s superior retrieval performance on three multi-hop question answering datasets. Additionally, our system achieves state-of-the-art results with improvements exceeding 10% on the challenging MuSiQue dataset, while requiring fewer tokens and iterations compared to other multi-step retrieval systems.