We're going to real time predictions

In this workshop, commuity member Jacopo Tagliabue, Director of AI of Coveo discusses their “Post-Modern Data Stack”, that is, a deconstruction of the MLOps stack they previously shared with the community.

In particular, they join the “modern data stack” (Snowflake + dbt) with “modern MLOps” practices using Metaflow to bridge the gap between data, training, and inference in a pure serverless fashion.

As usual, Jacopo refuses to work in a toy stack and with toy data: leveraging Coveo's huge data release from last year, they walk through a real-world recommendation pipeline, going from raw data to a live endpoint serving predictions.

Download the "modern data stack with the modern ML stack" here. be prepared to code along with us in today's meetup! Happening at 5pm BST/ 9am PST/ 12pm EST. Jump in!

Building recommender systems, one step at a time. David Hershey takes us through the architecture and design of three different types of movie recommenders.

First loading movie recommendations for users at the end of each day using historical data. Also known as the batch approach.

Second surfacing gems in the "you might also like" recommendations after watching a movie in real time, using historical data. Aka, semi real time.

And lastly, recommending movies in real time after user enters a search query, using fresh, real time data. Aka.... I'll let you guess

Props go out to David and his presenting style, he made sure to not distract us with flashy design and stuck focusing what matters most!

Yash Sheth, co-founder and VP of Engineering at Galileo came on the pod to talk about some key challenges in the unstructured data realm.

So what are some of the main challenges when working with unstructured data?

Understanding your data/curating your data set - It's cumbersome. You can't just sample data across the labels you want to identify. Unstructured data contains more nuances like the sources the data comes from or modalities of the data (think accents in speech).
Understanding where the model makes mistakes - How can you debug something you don't understand or have visibility into? Yash talks about how he has seen MLEs using ad-hock scripts and looking at data dumps individually to try and understand why a sample is bad.
Changes Fast - Going back to that speech example, think about how often the words we use to refer to something change. Back in my day, they called it reefer... now all the cool kids call it ZaZaaaaa.

Last Thursday on June 9, we had an #ask-me-anything session with @Neal Lathia, director of ML at Monzo Bank in the UK, ex-Skyscanner, PhD and creator of open-source library Modelstore for model versioning.

Q (Shri Javadekar): What metrics do companies track for the efficiency of MLOps?
A (Neal Lathia): We look at how long it takes us to ship, how long the feedback loop is for that specific problem, and how much undifferentiated work was required.

Q (Lina Weichbrodt): What are the biggest non-technical challenges you faced and what worked best to solve them?
A (Neal Lathia): The biggest non-technical challenge is when the application of ML occurs in an area that requires formulating the problem in a way that domain experts don't usually formulate it.

Q (Bozhao Yu): Can you tell us the CI/CD pipeline and testings?
A (Neal Lathia): It varies across our different stacks. This opens up the deeply philosophical question of what the purpose of unit tests are and whether they're the right thing to use to gain sufficient confidence that something works.

Deploying a model is only the first step. Come join some of top platforms dedicated to monitoring, observability and explainability to get the answers you need to keep those models running smoothly in production.

Destroy drift. Track down unexplained anomalies. Stop inference latency in its tracks.

Hear from companies laser-focused on machine learning production: Arize AI, Fiddler, WhyLabs, TruEra, Aporia, Bosch AIShield, Iguazio, ClearML, Modzy, InfuseAI, Superwise, Seldon, Toloka.

Get registered right here with just a few clicks.

Inspired by Ben Kuhn’s Essays on programming I think about a lot, Vishnu put together a list of the most influential reads on his journey as a machine learning engineer. Highly recommend all of these essays as must-reads for MLEs at any stage of their career.

There are timely reads and timeless reads. Timely reads are knowledge; timeless reads are wisdom. Throughout his MLE career, he's been bombarded with information and knowledge, much of which has been merely timely. Vishnu doesn't consider these to be worth re-reading and truly absorbing.

This list is heavily biased towards timeless reads. The lessons they contain are wise enough to stand up for years (as many of them have). The insights you can apply to elevate your approach to work over the long-term are contained in these timeless reads.