# ML Platform Tradeoffs and Wondering Why to Use Them

Javier Mansilla, Mercado Libre | MLOps Coffee Sessions | Episode 88 | 53:57
Hosted by Demetrios Brinkmann

Source: https://www.youtube.com/watch?v=DQM-Jue-QiE
Channel: MLOps Community, now AAIF Live (https://www.youtube.com/@AAIFLive-x1r). Summarised by MLOps Talks.
Page: https://mlopstalks.com/talks/ml-platform-tradeoffs-and-wondering-why-to-use-them
Published: 2022-03-28
Tags: build-vs-buy, deployment, platform-teams, team-adoption

## TL;DR
- Javier Mansilla says Mercado Libre builds shared ML foundations so teams can move quickly without each owning monitoring, deployment, data safety, and other operational work.
- Mercado Libre measures its platform partly by how many ordinary technology teams start using machine learning, then looks at adoption and business value across individual units.
- Javier's experience with AutoML is that reusable domain knowledge and understanding of a dataset are difficult to beat with automated model search.

## Summary
Javier Mansilla describes Mercado Libre as an e-commerce, fintech, advertising, credit, and logistics ecosystem serving Latin America. Its ML platform team builds shared foundations for internal teams rather than forcing all model development into one central group. The reason is operational: teams can deploy quickly, but monitoring, redeployment, data safety, governance, and provider changes still need ownership. Mercado Libre also uses several cloud providers, so its platform abstracts infrastructure where possible and lets teams move models without changing application code. Javier explains how the company encouraged adoption through workshops, boot camps, and small MVP teams that tested whether a problem could create business value. He is candid about trade-offs. The platform must support beginners while leaving room for experienced users to customize pipelines and artifacts. The team is mostly made up of engineers with a product mindset, working alongside data scientists. Javier's AutoML example ends with a preference for capturing expert knowledge and reusing it across teams.

## Key ideas
### Mercado Libre builds shared foundations instead of centralizing every ML team
[06:48](https://www.youtube.com/watch?v=DQM-Jue-QiE&t=408s)
Javier explains that his team does not take machine learning away from business units. It builds the foundations so each team does not have to reinvent deployment, monitoring, redeployment, and data-safety work. Mercado Libre learned this lesson from teams leaving the company, technologies becoming deprecated, and vendor lock-in. The platform is part of a wider company approach that also covers web development, APIs, and databases. Business units can still build their own first versions when the central team is not ready, especially in areas such as marketing and advertising. Those local experiments can later become services for the rest of the company.

### A platform becomes worthwhile when company scale makes repeated operational work expensive
[11:30](https://www.youtube.com/watch?v=DQM-Jue-QiE&t=690s)
Javier contrasts Mercado Libre with a startup of roughly 60 or 70 people, where building a dedicated platform would not have made sense. At Mercado Libre, the platform had a team of about 40 to 50 people, alongside a separate team of about 15 working on A/B testing. The company had thousands of technology workers and was continuing to grow. That scale made shared infrastructure useful, even though individual developers sometimes wanted to try tools the platform did not provide. Javier presents platform building as a trade-off that depends on the company's size, resources, and timing.

### Multi-cloud requirements drive Mercado Libre to maintain its own abstraction layer
[12:53](https://www.youtube.com/watch?v=DQM-Jue-QiE&t=773s)
Javier says Mercado Libre chose to use multiple cloud providers to reduce dependence on commercial or infrastructure problems at a single provider or location. Its ML platform sits on another platform that abstracts the underlying cloud services. Teams interact with Mercado Libre's internal applications instead of directly calling a provider service such as Amazon SageMaker. This lets the platform move a deployed model or service to another provider without requiring application code changes. Javier says the company had more than 500 deployed models, so changing each one by hand would not be practical.

### Platform adoption starts with culture change and measures who is using machine learning
[17:18](https://www.youtube.com/watch?v=DQM-Jue-QiE&t=1038s)
Mercado Libre initially tracked the percentage of people on regular IT teams who began experimenting with machine learning. Javier describes this as the company's machine-learning market share. Later, the platform team began examining adoption across individual business units, since advertising could be using machine learning heavily while shipping or other parts of the business used it less. The team also looks at platform adoption because duplicated work and personally identifiable information risks are easier to manage through shared systems. Javier says the broader goal is useful business value delivered quickly, rather than simply increasing request volume.

### A useful platform has to lower the barrier for beginners without blocking experts
[24:10](https://www.youtube.com/watch?v=DQM-Jue-QiE&t=1450s)
Javier says Mercado Libre tries to support beginners and experts at the same time. Beginners need a simple entry point, while experienced users need to inspect what happened and change fine-grained settings. The platform originally offered rigid entry points for dataset creation and model training. That design did not fit complex solutions built by hundreds of teams. The team moved toward checkpoints and artifacts, helping users check artifact quality while allowing them to build different kinds of pipelines. Users can create models that produce datasets or models that feed other models. Javier says the platform team is now paying more attention to experts who want permission to change or replace individual building blocks.

### Workshops and MVP teams teach business groups what machine learning can and cannot do
[29:01](https://www.youtube.com/watch?v=DQM-Jue-QiE&t=1741s)
Mercado Libre used workshops for business and product groups in areas such as logistics, fashion, payments, and its e-wallet. The sessions showed recurring patterns, such as tasks handled by rule engines or repeatable tasks where machine learning might help. Participants then worked with a small dataset to label potentially counterfeit products. The exercise made the limits visible. A model could provide a probabilistic answer, but it could not act as a perfect true-or-false oracle. Later, MVP teams worked with business groups for a limited period to produce a rough solution and test whether there was real economic value. Javier says these teams helped break down silos and persuade groups to try machine learning.

### The platform team is mostly engineering-led and works with data scientists on product decisions
[37:52](https://www.youtube.com/watch?v=DQM-Jue-QiE&t=2272s)
Javier argues that strong engineering teams are needed to build an ML platform. He names quality, reliability, and repeatability as areas where engineering experience matters. His own team is mostly made up of engineers who have built Python or web applications, even when they have not previously built ML systems. A neighboring team led by Rafael includes data scientists and engineers who work on data science. The teams define what to build together. Javier also says the platform needs a product mindset because its users include ML teams and the company itself. A new feature has to be weighed against its reliability and possible business impact.

### AutoML works better when it reuses internal expertise than when it searches blindly
[50:06](https://www.youtube.com/watch?v=DQM-Jue-QiE&t=3006s)
Javier describes Mercado Libre's AutoML effort, called AutoMeli, as a place where a team can send a dataset and receive a model. The team first uses internal cookbooks based on patterns it has seen in its own data. Javier's point is that a person who has spent a year studying a dataset may know which model family is likely to work well. AutoMeli can preserve that knowledge and make it available to other teams working with the same data. In his experience, automated machine learning has been hard to beat when compared with domain knowledge and a strong understanding of the data.

## Notable quotes
- Javier Mansilla: "Instead of centralizing, we try to provide those blocks." (08:00)
- Javier Mansilla: "We do have the power to make it appealing, and it is something we try to do." (21:03)
- Javier Mansilla: "We need to have a product mindset." (41:03)
- Javier Mansilla: "It is very hard for us to see those automatic machine learnings beating the domain knowledge." (53:41)

## Tools & references mentioned
- Mercado Libre
- Mercado Pago
- Fury Data Apps
- Amazon Web Services
- BigQuery
- Python
- AutoMeli

## Who should watch
- You are deciding whether an ML platform is justified for your company's size and internal team structure.
- Your platform team needs to balance simple workflows for new users with customization for experienced engineers.
- You are evaluating AutoML and want to compare automated search with reusable knowledge about your data and business domain.

## Editor's note

Javier Mansilla says Mercado Libre built a shared platform because repeated deployment, monitoring, redeployment, and data-safety work became expensive across its teams. ZenML records each pipeline run's steps, inputs, outputs, and code version, so teams can trace a model or artifact back to the data and code that produced it. Its configurable stack also lets the same pipeline run across different infrastructure.

Written by the MLOps Talks editors (the ZenML team), not by the speaker.

## Related talks

- [Machine Learning at Scale in Mercado Libre](https://mlopstalks.com/talks/machine-learning-at-scale-in-mercado-libre) (Carlos de la Torre, Mercado Libre, 59:28)
- [ML Platforms, Where to Start?](https://mlopstalks.com/talks/ml-platforms-where-to-start) (Olalekan Elesin, HRS Product Solutions GmbH, 52:52)
- [Build vs Buy an ML Platform](https://mlopstalks.com/talks/build-vs-buy-an-ml-platform) (Diego Oppenheimer, Algorithmia, 57:20)
- [The Shipyard: Lessons Learned While Building an ML Platform](https://mlopstalks.com/talks/the-shipyard-lessons-learned-while-building-an-ml-platform) (Joseph Haaga, Interos, 40:00)
- [Platform Thinking: A Lemonade Case Study](https://mlopstalks.com/talks/platform-thinking-a-lemonade-case-study) (Orr Shilon, Lemonade, 51:42)