Aitor Chamizo

I have worked on the implementation and investigation of different methods over heterogeneous data to perform feature selection and evaluate its mathematical stability, its ability to develop more interpretable models, and its impact on the environment. I have also developed state-of-the-art model wrappers and aggregators for Support Vector Machines, Gradient Boosting, and Random Forest, and I have contributed to the development of Logistic Regressor model wrapper in a Federated Learning scenario. I developed from scratch a model wrapper and a model aggregator for Large Language Models, compatible both for GPU and CPU, to allow the fine-tuning and inference of LLMs in the context of resource-limited hardware. I have worked in the problem space of asymmetric matching, regarding Natural Language Processing, where I developed a semantic matcher from scratch that improved a 20% both in terms of computational complexity and predictive performance, the matcher that the global company uses and the matcher that was put in production by previous colleagues for our local team. I developed from scratch a model factory to facilitate the process of time-series analysis over multiple products in pharmaceutical markets. This model factory allowed to orchestrate the management of artifacts, using Optuna to auto-optimize both naive models and linear-econometric models, whilst providing a road for analysis of results, metrics and figures to deliver the best results possible for the clients.

I was assigned the research line of Large Language Models, combining this work with my Data Science responsibilities. I developed from scratch a model wrapper and aggregator for Federated Learning applied to text-to-text Large Language Models (LLM), being able to train on both CPU and GPU in resource-limited devices; this allowed Sherpa.ai to position itself as a competitor in the LLMs niche. I also worked on the implementation of NLP metrics to measure the predictive performance of the LLM and help develop a research line based in Federated LoRa for heterogeneous scenarios. I wrote several technical reports regarding use cases in Horizontal and Vertical

Federated Learning, in particular in the fields of Predictive Maintenance (time series research using between 5 and 100 nodes), Object Detection (for defense scenarios), and Image Classification (for collagen mutation diseases - NIH/UCL).

I have led and developed a collagen mutation project over microscopy images on a project made in collaboration with the NIH and UCL, customer churn projects in collaboration with Iberdrola, Telefonica, and Prosegur, object detection projects in collaboration with Leonardo and Indra, and cybersecurity threat detection in collaboration with NetApp, from the point of view of Federated Learning. I have performed an analysis of the competence from a technical point of view of other platforms to see improvements and study our position in the market and developed several benchmarks. I have developed state-of-the-art ways of training and aggregating parameters for Random Forests, Gradient Boosting, and Support Vector Machines models. I have also contributed to the development of the Logistic Regression model wrapper.