The platform team
Who builds the platform data scientists use, and how big should that team be? Then comes the recurring discovery that building it does not make anyone use it.
Running a Fintech on Machine Learning
Pushed backThe team favored building its feature store internally rather than buying an existing product because it wanted tighter integration and customization, and found no sufficiently mature market solution.46:45
Build vs Buy an ML Platform
Pushed backDiego Oppenheimer rejects the assumption that buying a platform necessarily removes flexibility, comparing it with the mix of purchased tools used in software development.40:57
Scaling ML Capabilities in Large Organizations
Pushed backJoe argues that organizations should avoid buying a model serving platform when a simpler, opinionated solution is sufficient.57:25
How to Choose the Right ML Tool
Pushed backMariya Davydova recommends building an internal MVP in the Feast situation, while the team is also considering waiting for the planned release or choosing another tool.33:31
2 tools = 90% operational ML
Pushed backThe panel distinguished Kubernetes as a common infrastructure substrate from Kubeflow as the machine-learning platform built on it.33:05
MLOps Engineering Labs Recap, Part 2
Pushed backLaszlo viewed direct Kubernetes use as too difficult for an average data scientist, while Artem argued that Kubernetes is highly useful and should be exposed through simpler platforms.39:23
MLOps Investments
Pushed backSarah Catanzaro does not favor end-to-end ML platforms as a general solution because they can be poorly suited to particular workflows, while point-product sprawl also creates maintenance and integration costs.30:24
Maturing Machine Learning in Enterprise
Pushed backKyle Gallatin rejects the idea that one platform can fit every machine learning use case and argues for integration-first platforms with room for customization.17:02
On Structuring an ML Platform 1 Pizza Team
Pushed backThe platform team's size was described as four technical people earlier and as five people later in the discussion.31:59
Platform Thinking: A Lemonade Case Study
Pushed backLemonade's platform does not require one fixed level of openness: experienced users can customize decisions while other users can rely on defaults.27:21
Feature Stores at Shopify and Skyscanner
Pushed backMatt Delacour says Shopify's internal library is a thin layer on top of Feast, rather than a copied subset of Feast.20:24
Lessons from Studying FAANG ML Systems
Pushed backErnest Chan argued that teams should not simply copy the priorities of large companies because each platform reflects its own company’s requirements.32:51
ML in Production: A DS from Ubisoft Perspective
Pushed backJean-Michel Daignan rejects the idea that data scientists should be required to adopt extensive unit and integration testing, emphasizing scalability testing instead.14:54
Airflow Sucks for MLOps
Pushed backStephen Bailey argues that data teams should balance business value and platform quality rather than always building the most rigorous solution before releasing anything.19:39
Machine Learning Education at Uber
Pushed backMelissa Barr said that a company without a homegrown machine learning platform may not need to build its own internal education program because strong external resources already exist.42:26
LLM on Kubernetes
Pushed backRahul Parundekar argues that companies should prioritize a repeatable deployment platform and model iteration over over-optimizing Kubernetes autoscaling while GPUs are scarce.16:42
AI Innovations: The Power of Feature Platforms
Pushed backNikhil Garg says a feature store and a feature platform are different: a feature store mainly provides storage and serving, while a feature platform also includes computation and other end-to-end capabilities.44:51
Global Feature Store: Optimizing Locally and Scaling Globally at Delivery Hero
Pushed backThe participants disputed whether one centralized machine learning platform should define tooling and priorities for all departments.4:03
Reinvent Yourself and Be Curious
Pushed backStefano Bosisio disputes the assumption that technically impressive ML platforms will be adopted automatically, arguing that internal communication and education are necessary.18:48
Building an ML Platform from scratch
Pushed backBen says Prefect is the one tool he believes is nearly always useful, while other platform tools can be overkill for a single person; Eric's discussion emphasizes starting with simpler monolithic pipelines.1:37:43
Re-Platforming Your Tech Stack
Pushed backThe team's machine learning operations requirements differed from what a platform team focused mainly on model development would initially provide.22:00
Comparing ZenML, Metaflow, and all the other DAG tools
Pushed backEric rejects the interpretation that the platform stack makes ZenML and Metaflow genuinely interchangeable without substantial work.29:49
Iceberg, MCP, and MLOps: Bridging the Gaps for Enterprise
Pushed backSimba Khadder and Caleb Baechtold explain Iceberg's stronger adoption than Delta by pointing to openness and weaker vendor lock-in, while noting Delta's Python support advantages.57:40
Hard Learned Lessons from Over a Decade in AI
Pushed backMike Del Balso disputes the assumption that platform teams can reliably summarize their value by adding up dollar estimates from internal users.32:26
Accelerating Growth Through Optimizing GPU Usage
ClaimAdobe's compute platform aims to improve developer productivity by simplifying access to GPU instances.3:43
From Notebooks to Production FASTER
ClaimThe team serves about 200 data scientists and develops platform tools with their feedback.2:01
A Playground for AI Engineers
Pushed backNewly discussed tools such as TüN should not be adopted merely because the industry is discussing them; they should solve a real problem first.40:30
How We Cut LLM Latency 70% With TensorRT in Production
Pushed backMaher Hanafi says senior engineers initially believed AI would not substantially change their work, but later adopted it heavily.48:45