Welcome to my website!

I’m Adrian Valente, currently Research Engineer at Instadeep in Paris, developing LLMs for genomic data. Previously, I was a PhD student in computational neuroscience, working under the direction of Srdjan Ostojic at ENS Paris, where I worked on explaining the dynamics of recurrent neural networks using low-rank approximations, and on applications to electrophysiological (notably prefrontal cortex) data. You can find my thesis here, as well as the list of publications here. I also worked at Centre Léon Bérard in Lyon with Stéphane Depil on models of epitope presentation for cancer immunotherapies, and at Microsoft in software engineering.

Here, I like to talk about neural networks (artificial and biological), and ML and software engineering for science.

Find me:

  • Low-rank RNNs: a summary of my PhD work on low-rank RNNs and their applications to interpretability of neural dynamics.
  • RNNs strike back: a post covering the field of linear RNNs, also known as state-space models, and how they came back in the race for SOTA sequence modelling when we thought Transformers had killed them.
  • Software engineering for scientists: a WIP of 2025, this is a book I am writing to help scientists master the best practices of software engineering, like version control, testing, CI/CD, containerization and many more.
  • hippoLLM: a project from 2024 exploring the use of hybrid graph-vector databases as long-term memory modules for LLMs.