Team Teardown: MLOps in the Cloud for IoT
summary
Built a Kubeflow Pipelines stack on Kubernetes with on‑prem GPUs, VS Code devcontainers and AWS Batch for synthetic jobs, supporting image models, RL, simulation and an embedded tf lite ecosystem runtime; conversion pipelines and a custom inference engine push models to robots. Data comes from a paid data collection fleet (opt‑in) and offline evaluation parallelizes across millions of images plus hardware‑in‑the‑loop tests. Onboarding is still hard, so internal tooling automates Docker/registry/versioning and enforces opinionated standards that largely work.
AI-written summary · about
chapters (18)
- 0:00Introducing the panel
- 0:42ML journey at iRobot
- 2:28Team composition and roles
- 4:10Building the ML platform
- 8:02Kubeflow and infrastructure challenges
- 10:28Operationalizing edge models
- 11:34Onboarding data scientists
- 14:02Collaboration between teams
- 15:47TFX and ecosystem choices
- 16:30Online learning and continuous training
- 17:55Custom inference engines
- 18:59Simulation in pipelines
- 20:00Data collection strategies
- 21:25Prioritizing platform needs
- 23:09Standardization vs flexibility
- 25:59Editor and tool standardization
- 26:59Keeping teams aligned
- 28:29Closing and Q&A
related talks
13mMLOps Pipelines to Deliver Tf.Lite Models to Distributed Compute Pools at the Edge
21mMastering the DevOps Machine Learning Pipeline
26mKubernetes and MLOps for Scalable and Reproducible Generative AI
33mBuilding & Managing MLOps Pipelines in the Cloud
22mBuilding a MLOps Platform with Kubeflow on GKE
49mKubernetes and MLOps for Scalable and Reproducible Generative AI
46mAccelerating Machine Learning DevOps with Kubeflow
20mML Ops design patterns with Kubeflow Pipelines