My name is Apo and I’m a mighty pirate

Educated in several acronyms across the globe (UNISR, SFI, MIT), I was co-founder and CTO of Tooso, a NLP / IR startup in San Francisco acquired by TSX:CVO.

I led Coveo’s A.I. and MLOps roadmap from scale-up to IPO, and built out Coveo Labs, an applied R&D practice rooted in collaboration (e.g. Stanford, Outerbounds, Netflix, Farfetch, Microsoft, NVIDIA), open source and open science.

I’m a proud NLP advisor for Plural, and Adj. Professor of ML at NYU, which is mostly notable because it is the only job I ever had that my parents (sort of) understand.

I recently started investing in tech startups, both directly and as LP in AI funds: I’m always happy to chat with founders about DataOps, MLOps and AI.

I’m now building Bauplan.

Where is my mind?

I occasionally share code, ideas and teaching materials; if you have no intention of selling me anything, you can also try me on Linkedin.

I talk a lot, and I’m sometimes invited to do so by friends in industry (e.g. Home Depot, Lowe’s, Farfetch, eBay, NVIDIA, Pinterest) and research (e.g. keynotes at KDD, SIRIP, RecSys): some of my talks and papers are highlighted at the end of this page.

Research-y stuff

My recent research sits at the intersection of language, learning and retrieval, with a recent drift towards data processing.

I am co-organizer of SIGIR eCom (2022, 2023), EvalRS (2022, 2023), Industry Sponsorship Chair for CIKM 2022, and I’ve been in the program committee for many conferences (COLING, ECONLP, ECNLP, EMNLP, ACL, SIRIP). My work has been presented in venues such as NAACL, WWW, RecSys and Nature journals: our paper on cognitively-inspired query embeddings won the Best Paper Award at NAACL 21.

As a true SFI alumnus, I am an old-fashioned generalist, and I gave tiny contributions to other fields mostly as a way to spend more time with old friends: data systems, logic and computation, cellular automata, computational social sciences, networks, philosophy of mind, political science.

Old stuff

In previous lives, I managed to get a Ph.D., simulate a pre-Columbian civilization, document biases in national elections and give an academic talk on videogames. Some of my improbable “achievements” received ample press coverage in national outlets.

Having built end-to-end data and ML pipelines at garage, growth and IPO scale, I happily shared all my mistakes in a series of articles that introduced the concept of Reasonable Scale.

Some time before Brad Pitt’s movie, I led one of the first attempts of running sophisticated analytics for a professional basketball team, and spearheaded the first data science effort on Milan’s bike-sharing service (no bikers or bureaucrats were harmed during the project).

About this page

The content of jacopotagliabue.it are released under the BY-NC-ND license; my chibi has been designed by the incredibly talented wisesnail.

Last update: August 2023.

Appendix

Quick links to some selected projects, talks, papers, datasets.

Open source projects

RecList is a testing library for recommender systems: RecList spawned a sponsored open source package, a competition, an open source hackathon, a white paper on evaluation and three articles (WWW, Nature Machine Intelligence, CIKM).
FashionCLIP is a fashion-aware model based on CLIP. As the first ever industry-aware CLIP model, FashionCLIP spawned an open source repo, two papers (Nat. Sci. Rep. and ACL), and a popular HuggingFace release, with >500k downloads in just the first few months.
“You don’t need a bigger boat” is an introduction to modern pipelines at the Reasonable Scale, first presented at the Stanford MLSys Lectures in 2021. A simpler setup (with the same principles) is my Post-Modern Data Stack.

Talks

Applied R&D at startup scale, SIGIR 2023 (SIRIP Keynote), 04/23
Wild Wild Tests, Arthur Ground Truth Series (Invited Talk), 04/23
Mo’ models, mo’ problems: CLIP-like models and generalization in modern e-commerce platforms, KDD 2022 (Ecom-Gen Keynote), 08/22
Are we there yet? Meaning in the age of large language models, CiE 2022 (Keynote), 07/22
Recs at reasonable scale (slides, repo), NVIDIA RecSys Summit (Keynote), 07/22
Not all those who browse are lost: few-shot and zero-shot personalization with deep architectures, Haystack, 07/20

Papers

A challenge for rounded evaluation of recommender systems (Nat. Mach. Int. 2023)
Contrastive language and vision learning of general fashion concepts (Nat. Sci. Rep. 2022)
Beyond NDCG: behavioral testing of recommender systems with RecList (WebConf 2022)
You Do Not Need a Bigger Boat: Recommendations at Reasonable Scale in a (Mostly) Serverless and Open Stack (RecSys 2021)
Query2Prod2Vec: Grounded Word Embeddings for eCommerce (NAACL 2021, Best Paper Award)
Language in a (Search) Box: Grounding Language Learning in Real-World Human-Machine Interaction (NAACL 2021)
Cellular Automata (SEP 2017)
There’s Plenty of Boole at the Bottom: A Reversible CA Against Information Entropy (Minds & Machines 2016)

Datasets and data challenges

EvalRS 2023, augmenting the EvalRS 2022 dataset with lyrics and sentiment information, plus a whole new set of tests. The dataset was used to organize the first-of-its-kind hackathon at KDD.
EvalRS 2022, packaging the LFM-1b dataset into an easy-to-use, “testable” format for our data challenge. The competition featured 150 participants in 50 teams, and was at the heart of a Nature Machine Intelligence editorial.
SIGIR eCom Challenge 2021, the largest and most complete session-based dataset of all time, released for SIGIR 2021.
Shopper Intent Prediction, from our paper.

Aside from research and tutorials, our datasets have been successfully used by dozens of master students to defend their thesis at Tillburg University and Politecnico in Milan.