Learn more about Jacopo on his linkedin.
I’m Jacopo, I’m Director of AI at Coveo. In my team, we combine product thinking and research-like curiosity to build better ML systems, with particular attention to eCommerce use cases (product search, recommender systems, personalization, intent prediction etc.).
We are huge fans of open science and open source, and divide our time equally between research, product and evangelization (mostly through code and datasets).
I have a life-long interest in (formal and natural) languages, and grew up as an old-fashioned generalist; I never took programming courses, and that explains why I still (kind of) suck even now!
In some sense, I have been doing “data science” before it was cool, or even a phrase; in other sense, I feel the field is changing so rapidly that my tooling and best practices keep changing as well.
Obviously, a huge boost in experience and confidence came from my entrepreneurial past, as founder and CTO of a NLP company in San Francisco, Tooso, which was acquired by Coveo in 2019.
Going from founder to manager in a much bigger company involved a significant shift in my day-to-day. I write less code now, and it is mostly for prototyping, pedagogical or research reasons: my job has shifted from “building out all sorts of stuff” to “asking which problems are worth solving”, and then mentoring teammates in actually solving them.
Finally, a significant part of my day is spent building our network of collaborations and strengthening the company’s position in the AI space. We actively evaluate ML tools of all types (tracking, monitoring, computing, etc.), and evangelize the community about our findings.
We are a B2B API company, so our product is, in some sense, ML models: we “know” that a problem is suited for ML because of our domain knowledge and product-market fit. That means that a lot of problems that traditionally plague “ML projects” in other industries - like data collection - are less of a burden for us, as we designed the system to be working end-to-end: no data, no product.
As a B2B company, we have two types of “users”: our client - say, a mid-to-large eCommerce -, which uses our APIs to train models and then provide its clients - i.e. the shopper - a better experience (better search ranking, personalized recommendations etc.). Hopefully the shopper is happier thanks to our models, and buys more from our client, which in turn is happy to invest even more in our solution, and so on.
That said, we come from years in the industry, having built true end-to-end systems in the space: back in Tooso days, it was not uncommon for me to be in a sale call, an investor meeting, a product retrospective and a coding session in the same day. Together with the ability to collaborate with PMs and less technical people, I truly believe that a startup experience makes for much better employees.
Well first things first: is it really a problem that requires ML? While ML as a field has improved massively in the last ten years, there are still many use cases for which ML is either a bad choice, or practically impossible (e.g. a model could be built, if only we had access to secret dataset XYZ).
If the problem requires ML, I do some Googling and literature research to find what others have tried before: in many cases, the solution is either out there, or at least, half there (and you can use existing methods as baselines).
I think an important concept, in general, is “knowing when to stop”: as we argued at length, most ML practitioners outside of Big Tech deal with problems at “reasonable scale”; in other words, the marginal gain of increasing model sophistication is often soon surpassed by the additional training / maintenance cost.
The team’s philosophy is the MLOPs with no Ops approach: we are obsessive about data quality and standardization, and we put a lot of work upfront to make sure new hires and collaborators can easily access data, train models, and share their results.
We maintain almost no infrastructure leveraging cloud-based solutions, and we require ML engineers to own the entire process, from data aggregation/filtering/preparation to testing and documentation. The job of people like me (and my counterpart in data engineering) is not doing X or Y, but making sure our team can do X or Y safely and autonomously.
I never worked at FAANG-scale, so my perspective is always at “reasonable scale”.
Teams that work are small (a casual observer would say “understaffed”), focused and own the work 100%: we subscribe to the “end-to-end ML practitioner” view, and encourage younger teammates to get curious about the use case, not just the tech.
On the other side, a common anti-pattern is hiring dozens of data scientists without first figuring out data, tooling and a roadmap: now you have a swarm of expensive employees doing notebooks that go nowhere and spending time moving data around. To all the people building up teams out there: hire very few key people first, picking experienced scientists that did it all _before; once they figure out data and tooling, _then hire ICs: you will find out that you need way fewer people if everybody is productive.
As our friend Ville says, “production-ready is a continuum”: this is one the wisest things out there regarding ML systems.
We use Metaflow for both research and production pipeline, SaaS products for experiment tracking (Comet, Wandb), and PaaS offering for quick serving (SageMaker): building a working prototype is much better than slides, and in today’s cloud offerings there are plenty of one-liners to set up an endpoint from a model artifact produced with Metaflow.
All our data is organized in Snowflake, and ready to use with pre-made tables abstracting away the complexity of the ingestion process.
ML for us starts always with data: as the Data-Centric AI folks rightfully highlight, data collection, cleaning, aggregation is a part of the ML system, as it conditions its behavior as much as (or more than) models. On the data side, we have been enthusiastic adopters of Snowflake, which retired Redshift, Athena and EMR Clusters and finally provided the PaaS experience everybody was waiting for. On top of that, we currently like DBT and Great Expectations for transformation and QA.
On the more proper ML part, we moved to Metaflow last year and thoroughly love the framework: its design philosophy is very much aligned with our view of ML. We use Metaflow both for research pipelines and product work.
In a narrow sense, we have target KPIs to monitor: internal, such as average click rank, and external, such as average order value for eCommerce sales. To have reliable results, it is important to devote significant effort in online testing and causal analysis: if I have to pick one, this is the most fundamentally overlooked and unsolved problem in the space.
In a broader sense, we are part of the AI community: we publish papers, open source code and release datasets to advance the community as a whole. There are some proxies - citation, invited talks, downloads - but we do more of a qualitative assessment: we organized the SIGIR 2021 eCom Challenge, we recently won NAACL Best Paper Award, and in 2020 Coveo was one of the most present companies in eCommerce top-tier research venues, next to giants like Amazon, Alibaba, Etsy.
Finally, in a much more pop definition of “impact”, there is the unique pride in building stuff that people use, even without knowing it: if we do some back of the envelope calculation, Tooso interacted with approximately 1 out of 6 adults in Italy, which is pretty badass for a company almost literally out of a garage. Now at Coveo, our models work at scale on more than 500 clients, including several Fortune 100 companies.
A lot of my favorite ML people are current or former entrepreneurs, or at least embody the end-to-end ownership of a problem that is the hallmark of entrepreneurship: there is no such thing as “that is somebody else’s job description”, and there is genuine curiosity (and, of course, matching talents) to understand all the parts of the system, as well as where the whole is going - why are we doing this in the first place?
Training is nothing, will is everything: it matters less what you already know and did, than what you’re willing to learn and do.
We have ties with other institutions and researchers: some old friends from our previous lives in academia, some new friends we picked up along the way. Working with curious researchers is not just stimulating, but also a fantastic way to keep learning new things directly from the people making the field: progress is better (and more fun) when shared.
Practically, I’m pretty bad with systematic methods to extend my knowledge, and I follow a curiosity-driven approach, picking from conferences, random browsing and friends’ suggestions my next thing to read. If I have to pick a foundational book in AI, Gödel, Escher, Bach is my obvious choice; if I have to pick a researcher to follow closely these days, Josh Tenenbaum; if you want to know what I’ve been reading recently (both delightful!), “Causal Inference: The Mixtape” by Scott Cunningham, and “Mostly Harmless Econometrics” by Joshua Angrist (before he was cool) and Jörn-Steffen Pischke.
On a more philosophical point, I grew up with the SFI attitude: the world is a complex object, and it is unlikely that any one field in isolation will be enough to explain non-trivial phenomena - so you better keep learning from yours and other fields as well (for example, I just took an econ course to better understand marketplaces). This is the way.
Read more mentor interviews?
© Eugene Yan 2023 • About • Suggest edits.