Machine Learning
The Hidden Cost of Training Data Quality
Bad labels and inconsistent data cost more than bad models do, because they're harder to detect and they cap your ceiling no matter how good the model gets.
3 min read
Tag
MLOps
8 articles
Machine Learning
ONNX and the Case for Model Portability
Why exporting models to ONNX solves real deployment problems, plus the conversion pitfalls that catch teams off guard the first time they try it.
3 min read
Machine Learning
Model Serving Architectures: Batch, Online, and Edge
A comparison of the three main ways to serve ML predictions in production, and how to pick the right one based on latency and data freshness needs.
3 min read
Machine Learning
A/B Testing Machine Learning Models in Production
Offline metrics tell you a new model is better in theory. A/B testing tells you whether it actually moves the numbers that matter — here's how to do it right.
3 min read
Machine Learning
Batch vs. Streaming Feature Computation
How to decide when features need real-time streaming computation versus a scheduled batch job, and how to keep the two consistent with each other.
3 min read
Machine Learning
Detecting and Handling Model Drift in Production
Why models degrade quietly after deployment, how to detect drift with statistical tests before it shows up in business metrics, and what to do about it.
3 min read
Machine Learning
MLOps Pipelines: From Notebook to Production
How to turn a working Jupyter notebook into a reproducible, monitored production pipeline without rewriting everything from scratch.
3 min read
Uber
Michelangelo: How Uber Turned Machine Learning Into a Self-Service Platform
How Uber built Michelangelo to give data scientists a shared path from feature data to a deployed, monitored production ML model.
3 min read