# MLOps Week 29: Building the Future of ML Platforms with Ketan Umare

Page: https://stenobird.com/podcast/mlops-weekly/mlops-week-29-building-the-future-of-ml-platforms-with-ketan-umare
Text version: https://stenobird.com/podcast/mlops-weekly/mlops-week-29-building-the-future-of-ml-platforms-with-ketan-umare.md
Podcast: [MLOps Weekly Podcast](https://stenobird.com/podcast/mlops-weekly)
Published: 2024-05-30T20:45:33+00:00
Episode link: https://rss.com/podcasts/mlops-weekly/1504403
Audio file: https://content.rss.com/episodes/132586/1504403/mlops-weekly/2024_05_30_20_45_07_f36d5a19-a5d3-4259-923c-ee102c69f6f9.mp3
Processing state: processed
JSON: https://stenobird.com/v1/public/podcasts/mlops-weekly/episodes/mlops-week-29-building-the-future-of-ml-platforms-with-ketan-umare
Duration seconds: 1893

## Resource

Ketan Umare discusses the evolution of machine learning infrastructure, from the early days of managing models at Lyft to the creation of Flyte and his new venture, Union. The conversation explores why traditional orchestrators like Airflow fail at scale and how the rise of LLMs is reshaping, rather than replacing, classical machine learning.

## Highlights
- Main idea: Machine learning products require continuous investment and evolution because they are transformative and subject to changing user expectations
- Failure mode: Using general-purpose orchestrators like Airflow for large-scale ML workloads leads to scaling and consistency issues
- Practical takeaway: The most cost-effective and reliable production systems often use an ensemble of LLMs for orchestration and classical ML for transactional tasks
- Main idea: The distinction between software engineering and data science is critical, as the experimentation-heavy nature of science requires different infrastructure
- Future outlook: LLMs are a powerful new tool in the belt, but they are unlikely to replace specialized deep learning or gradient-boosted models for tasks like fraud detection

## Topics

MLOps, Machine Learning Infrastructure, Large Language Models, Flyte, Data Science Engineering, Cloud Computing, AI Orchestration, Model Deployment

## Chapters
- 1:00 — Background and the Lyft Era: Ketan shares his journey from high-frequency trading to leading the ETA prediction models team at Lyft.
- 3:20 — The Birth of Flyte: The transition from manual model deployment to building a scalable, operationalized ML workflow platform.
- 8:05 — The Limitations of Airflow: An analysis of why traditional data orchestration tools lack the necessary features for modern ML experimentation and scaling.
- 17:05 — The Impact of LLMs on Software: How the accessibility of OpenAI's models is forcing legacy companies to rethink their product interfaces and intelligence.
- 24:25 — Managing the Cost of AI Experimentation: Addressing the massive computational expenses associated with running large-scale multi-armed bandit tests and chatbots.
- 28:50 — The Future of the ML Stack: A discussion on why classical ML models like XGBoost will remain dominant in production alongside generative AI.

## Actions

- request_transcript: `POST https://stenobird.com/v1/public/podcasts/mlops-weekly/episodes/mlops-week-29-building-the-future-of-ml-platforms-with-ketan-umare/transcription-requests` — Idempotently request low-priority transcript generation for this episode.
- read_markdown: `GET https://stenobird.com/podcast/mlops-weekly/mlops-week-29-building-the-future-of-ml-platforms-with-ketan-umare.md` — Read the agent-friendly Markdown representation of this episode resource.

A page view does not enqueue transcription. Agents should invoke `request_transcript` explicitly when they need this episode processed.

## Transcript

Full transcripts are not published on public pages unless there is a clear rights basis.