# ML platforms people actually use

A pack of 8 sessions from the MLOps Community YouTube channel, in the order to watch them. 6h 03m of video.
Page: https://mlopstalks.com/packs/ml-platforms-people-actually-use

You built the platform, but data scientists still run their own scripts and ask colleagues how to deploy. Another capability will not help until the supported route makes their work easier. Start with Scout24's interviews and notebook prototype, then Netflix's distinction between removing operational work and hiding useful information. Lemonade and Mercado Libre show why convenient defaults need room for experienced users, and why a platform need not absorb every useful service. Interos makes standards actionable instead of another support ticket. Wolt turns documentation and user contributions into part of the product. Finish at Stitch Fix with evidence of adoption and pain removed. These accounts date from 2020 to 2023; their value is the reasoning behind the interfaces, rather than a shopping list of tools.

## This pack is for you if

- Data scientists keep using homegrown scripts even though your platform supports their workflow.
- Your standard path works for new users, but experienced users keep asking to bypass it.
- Your platform has shipped plenty of features, but you cannot show whose work got easier.

## The talks, in order

### 1. ML Platforms, Where to Start?

Olalekan Elesin, HRS Product Solutions GmbH | 52:52 | MLOps Coffee Sessions
Video: https://www.youtube.com/watch?v=HuSVC5XKOaI
Summary: https://mlopstalks.com/talks/ml-platforms-where-to-start.md

Why first: Elesin starts with people who were already working around slow notebook environments. Interviews and a working prototype gave them a reason to switch. Scout24's experience establishes the first job: discover which work users actually want removed before deciding what the platform should hide.

### 2. Human-centric ML Infrastructure: A Netflix Original

Savin Goyal, Netflix | 56:07 | MLOps Meetup
Video: https://www.youtube.com/watch?v=TzRNZO2E-eM
Summary: https://mlopstalks.com/talks/human-centric-ml-infrastructure-a-netflix-original.md

Why second: Removing tedious steps should not remove choices data scientists need to keep. In Goyal's Netflix account, users choose their modeling tools while Metaflow removes container and data-transfer work. It still exposes what happened in a run, so the simpler workflow remains possible to debug.

### 3. Platform Thinking: A Lemonade Case Study

Orr Shilon, Lemonade | 51:42 | MLOps Coffee Sessions
Video: https://www.youtube.com/watch?v=KemCHs7Xbrs
Summary: https://mlopstalks.com/talks/platform-thinking-a-lemonade-case-study.md

Why here: Lemonade narrowed its framework too far. Shilon describes the move from an open interface to a restricted one, then support for both, with extra responsibility for custom work. This tests Goyal's argument for freedom against an actual redesign. A Slack bot also puts routine operations where users already work.

### 4. ML Platform Tradeoffs and Wondering Why to Use Them

Javier Mansilla, Mercado Libre | 53:57 | MLOps Coffee Sessions
Video: https://www.youtube.com/watch?v=DQM-Jue-QiE
Summary: https://mlopstalks.com/talks/ml-platform-tradeoffs-and-wondering-why-to-use-them.md

Why here: Mercado Libre found another way to allow different workflows. Mansilla describes moving from rigid dataset and training entry points toward checkpoints and artifacts that different pipelines could share. After Lemonade's custom route, this offers a different boundary: agree on what needs checking without prescribing every step a user takes.

### 5. Fury Platform and Fury Data Apps at Mercado Libre

Carlos de la Torre, Mercado Libre | 10:30 | MLOps Meetup
Video: https://www.youtube.com/watch?v=d-f-2zPCRUA
Summary: https://mlopstalks.com/talks/fury-platform-and-fury-data-apps-at-mercado-libre.md

Why here: A platform need not absorb every useful service. De la Torre's earlier Mercado Libre account shows how Fury Data Apps could use another team's data infrastructure without taking it over. The flexibility in Mansilla's account extends here from individual workflows to services that other teams already run.

### 6. The Shipyard: Lessons Learned While Building an ML Platform

Joseph Haaga, Interos | 40:00 | MLOps Coffee Sessions
Video: https://www.youtube.com/watch?v=_n-GthGQw7o
Summary: https://mlopstalks.com/talks/the-shipyard-lessons-learned-while-building-an-ml-platform.md

Why here: An interface check is only helpful if the user can act on its failure. Haaga describes Interos automating its standards while treating confusing errors and unnecessary rules as platform problems. After connecting services through common interfaces, the practical test is whether an engineer can correct a mistake without waiting for the platform team.

### 7. Building an ML Platform: Insights, Community, and Advocacy

Stephen Batifol, Wolt | 45:49 | MLOps Podcast
Video: https://www.youtube.com/watch?v=mlrs8pe9QLI
Summary: https://mlopstalks.com/talks/building-an-ml-platform-insights-community-and-advocacy.md

Why here: Self-service needs continuing work after the interface ships. Batifol describes Wolt's getting-started guides, regular feedback and users sharing examples. Inviting users to contribute fixes extends Haaga's aim of helping people unblock themselves, so the platform team is not the only place to find an answer.

### 8. Aggressively Helpful Platform Teams

Stefan Krawczyk, Stitch Fix | 51:52 | MLOps Coffee Sessions
Video: https://www.youtube.com/watch?v=az8lXG9v4uo
Summary: https://mlopstalks.com/talks/aggressively-helpful-platform-teams.md

Why last: Krawczyk asks what evidence adoption work leaves. Stitch Fix built abstractions with an initial partner team and watched team adoption, deployed services and user feedback. His difficulty isolating time saved is useful restraint after Wolt's account of community work: a growing model count alone cannot show whether the platform removed anyone's pain.

## Editor's note

Goyal argues that hiding infrastructure should not hide what happened in a run. Shilon wants a consistent starting point without ruling out custom work. ZenML lets teams define workflows as Python steps and run them on different infrastructure through a configurable stack. Each run records its steps, inputs, outputs and code version, so a shared interface can still leave a record engineers can inspect.

Written by the MLOps Talks editors (the ZenML team), not by any of the speakers.