In this episode of MLOps Live, Sabine and Stephen are joined by Adam Sroka, Head of Machine Learning Engineering at Origami. They explore principles and frameworks for creating a team culture in MLOps that prioritizes the most important things and sets the team up for success. Collaboration between teams is necessary to move through the ML life cycle as rapidly and effectively as possible. Adam gives us an idea of what the MLOps culture is at Origami, how he built it, the challenges he encountered, and actions teams can begin taking to build a good MLOps culture. He identifies areas that businesses may capitalize on, including infrastructure, team structure, tools, project ownership, and KPIs for efficient workflow. Adam has provided clear insights into the methods and tools he has used to build great teams with good quality MLOps culture built over his career. Adam is excited to share his technical and non-technical expertise, and shed some light on what works, especially now that best practices and playbooks for how to build a good MLOps culture and maximize the value from your projects and teams are not yet readily available.

In this episode of MLOps Live, Sabine and Stephen are joined by Andy McMahon, Machine Learning Engineering Lead of the NatWest Group. They explore concepts around building your first MLOps systems and how teams can understand the processes of optimizing level 0 operations and move towards scalability. As soon as you commit a piece of code, a properly mature MLOps pipeline may be so powerful that it may be put into production immediately. However, attaining this level of maturity is extremely uncommon. Therefore, it becomes crucial to outline the requirements for creating a system that simplifies future operations, lowers failed deployments, and boosts performance. The goal is to build an MLOps system that you can easily iterate on and would not break when the time for scale and integrating components (such as model registry and feature stores) arrive. Andy demonstrates how a basic model with an optimal MLOps infrastructure will yield value more quickly than a complex model that is thrown over the fence, which may result in resource wastage. Teams can begin by redefining the deliverable expectations, simplifying them to what is truly necessary, utilizing available tools, and constantly realigning operational considerations to the business problem to be solved. Andy outlines several essential ideas, from the most basic level (MLOps at level 0), which includes no automation, to the most advanced one (MLOps level 1 and 2), which involves automating both machine learning and CI/CD pipelines.

In this episode of the MLOps Live podcast, Mateusz Opala, Senior Machine Learning Engineer at Brainly, answers questions about leveraging unlabeled image data with self-supervised learning or pseudo labeling.

In this episode of the MLOps Live podcast, Michal Tadeusiak, the Director of AI at deepsense.ai, will be answering questions about managing computer vision projects, specifically looking at AI activities spanning computer vision, NLP, and predictive modeling projects.

Today, we’re joined by Fernando Rejon, Senior Infrastructure Engineer at Zeta Alpha Vector, and Jakub Zavrel, Founder and CEO of Zeta Alpha Vector. In addition, they discuss MLOps for neural search applications data engineering, and how this innovation is pushing the bounds of search engines. In this episode, they explore how they use modern deep learning techniques to build an AI research navigator at Zeta Alpha. They engage in an in-depth discussion based on the challenges with setting up MLOps systems for neural search applications, how to evaluate the quality of embedding-based retrieval, progress and numerous pertinent criteria, contrasting the trade-offs of using in neural (information retrieval) search, and the trade-off with using it in practice and theory to standard information retrieval strategies. Additionally, they put into perspective the most important components you would need to build a POC neural search application. examine neural search models in both the retrieval and ranking phases from the perspective of scalability and predictability. They also outline conditions under which state-of-the-art results can be obtained. They also discuss the enormous work necessary to build and deploy neural search applications, which necessitates the use of greater processing resources, such as GPUs rather than CPUs, to get desirable output.

In this episode of MLOps Live, Sabine and Stephen are joined by Danny Leybzon, MLOps Architect at WhyLabs. They examine the differences between monitoring and observability in machine learning models for production and methods for efficient implementation and development. Observability in MLOps is a holistic and comprehensive way to gain insights into the behavior, data, and performance of a machine learning model throughout its lifespan. It allows for detailed root cause analysis of ML model predictions and aids in the development of responsible models. Although ML monitoring and observability appear to be similar, Danny points out that monitoring is a continuous system that prompts you when there is a problem. Whereas observability refers to the larger picture, a human-in-the-loop root cause analysis system that allows you to figure out what the problem is and then solve it. Danny further discusses the unique features of WhyLabs in comparison to other conventional monitoring solutions, such as customizable and opinionated self-serve capabilities that allow users to pick particular metrics to track, especially in the absence of ground truth.

Today, we’re joined by Federico Bianchi, a Postdoctoral Researcher at Università Bocconi. He discusses testing recommender systems, the essential features for any platform with that purpose, testing the relevance of these systems, and how to handle the biases they generate. With the continuous growth of e-commerce and online media in recent years, there are an increasing number of software-as-a-service recommender systems (RSs) accessible today. Users can get new content from recommender systems, which range from news articles (Google News, Yahoo News) to series and flicks (Netflix, Disney+, Prime Videos), and even products (Amazon, eBay). Today, there are so many products and information available on the internet that no single viewer can possibly see everything that is offered. This is where recommendations come in, allowing products and information to be classified according to their expected relevance to the user's preferences. They compared offline recommendations to online evaluation platforms, which allow researchers to evaluate their systems in live, real-time scenarios with real people. Federico discusses the benefits of offline modeling and evaluates the speed and convenience of testing algorithms with predetermined datasets. However, because these statistics are not tied to actual users, there are a lot of biases to consider.

In this episode of MLOps Live, Sabine and Stephen are joined by Kyle Morris, Co-Founder of Banana ML. They discuss running ML in production leveraging GPUs. They delve into GPU performance optimization, approaches, infrastructural and memory implications as well as other cases. With the increased interest in building production-ready, end-to-end ML pipelines, there’s an increasing need to employ the optimal toolset, which can scale quicker. Modern commodity PCs have a multi-core CPU and at least one GPU, resulting in a low-cost, easily accessible heterogeneous environment for high-performance computing, but due to physical constraints, hardware development now results in greater parallelism rather than improved performance for sequential algorithms. Machine Learning Build/Train and Production Execution frequently employ disparate controls, management, run time platforms, and sometimes languages. As a result, understanding the hardware on which one is running is critical in order to take advantage of any optimization that is feasible.

Today, we’re joined by Jacopo Tagliabue, Director of A.I. at Coveo. He currently combines product thinking and research-like curiosity to build better data-driven systems at scale. They examine how immature data pipelines are impeding a substantial part of industry practitioners from profiting from the latest ML research. People from super-advanced, hyperscale companies come up with the majority of ideas for machine learning best practices and tools, examples are Big Tech companies like Google, Uber, and Airbnb, with sophisticated ML infrastructure to handle their petabytes of data. However, 98% of businesses aren't using machine learning in production at hyperscale but rather on a smaller (reasonable) scale. Jacopo discusses how businesses may get started with machine learning at a modest size. Most of these organizations are early adopters of machine learning, and with their good sized proprietary datasets they can also reap the benefits of ML without requiring all of the super-advanced hyper-real-time infrastructure.

Today, we’re joined by Kuba Cieslik, CEO and Ai Engineer at tuul.ai. He has experience in building ML products and solutions and has a deep understanding of how to build visual search solutions. Visual search technology has been around for quite some time, as part of Google Pictures or Pinterest Lens. It has become increasingly popular in e-commerce, allowing customers to simply upload what they're looking for instead of going through a slew of attribute filters. Kuba discusses how one might go about creating such a visual search engine from the ground up, as well as what approaches work and the challenges in such a complex sector.