Filippo Pompili

Background

About Filippo Pompili

I have a PhD in Computer Science and 10+ years of industry R&D experience in Natural Language Processing (NLP), Information Retrieval (IR), and Machine Learning (ML). I worked in companies of all sizes: fast-paced startups, established large multinationals, and medium-sized companies, both in Canada and in Italy. Typical domains I've been working on are: question answering, information extraction, text classification, sparse and dense retrieval. I'm familiar both with the latest approaches (deep learning, LLMs, RAG, vector search.) and with traditional machine learning algorithms (max-margin classifiers, gradient boosted trees.). I know the best practices of data science workflows, including data exploration, problem modelling, algorithm selection, and feature engineering. Some of the technologies I use on a regular basis are: python, pandas, pytorch, huggingface's transformers, spacy, nltk, langchain, scikit-learn, xgboost, numpy, scipy, matplotlib, seaborn, shap, lime, jupyter lab, Elasticsearch, Amazon Sagemaker, Google Vertex AI and BigQuery, docker, SQL, FastAPI, poetry, git. I'm a strong advocate of the best practices for software engineering, including maintenance of a shared handbook for documentation, attention to experiments reproducibility, and adherence to git workflows standards. I code on a versioned and dockerized IDE stack which is easily deployable anywhere there's a terminal (completely without GUI); it's made of: alacritty + zsh + tmux + neovim. My PhD was focused around a powerful unsupervised learning algorithm: Nonnegative Matrix Factorization (NMF). For some text clustering and topic analysis applications, NMF is similar to Latent Dirichlet Allocation (LDA), but it has some peculiarities, e.g, computational efficiency, stability w.r.t. initial conditions, amenability to algorithmic integration of side information or domain knowledge, and, in some cases, much better interpretability (thanks to its sparse, parts-based, additive representations). It's been applied across many diverse scientific domains (e.g, dimensionality reduction, recommender systems, graph community detection, genetic data analysis, blind hyperspectral unmixing.). From this experience I have gained strong exposure to concepts from numerical optimization. Some of the principles I strive to abide by are the following ones: integrity, accountability, ownership, initiative, collaboration, humility, kindness, empathy, curiosity, ambition. I appreciate behaviour that aligns with them also from people, colleagues, and stakeholders, I work with.