Notes on data, AI, IT
and security
No marketing fog. The way I think about real problems with founders and managers.
The real cost of an NLP pipeline before you are sold by the demo
What actually requires ongoing support in a production NLP system - from data labelling to quality control in live operation.
Model drift: why an ML system degrades without visible failures
Machine learning models in production lose accuracy over time - quietly, with no errors and no alerts. What drift is and how to monitor for it.
From hype to inference cost: why AI must be measured as a production function
How to move from evaluating AI by its demo effect to evaluating it by the real economics of running a model in production.
Why AI projects die before they produce results: five recurring patterns
An analysis of the typical reasons AI initiatives stall or fail to deliver their promised impact - and what to do about it.
ML models decay silently - and most companies do not notice
A model that was accurate at launch will gradually stop being accurate as the world changes. Why monitoring for model decay is not optional, and how to set it up before it becomes an incident.
BERT and the new baseline for applied NLP
What the BERT model changes in the practical use of text processing, and why it matters for companies working with unstructured data.
Narrow AI in production: where the line between pilot and working system is
Why most AI pilots never reach production, and what it actually takes for a model to work in real conditions rather than just in a demo.
AI readiness: what companies confuse with actual preparation
Why the gap between interest in AI and operational readiness to deploy it is much larger than it appears after a conference or a demo.
Why ML teams keep rebuilding the same data pipelines
The hidden cost of ML at scale is not the models - it is the duplicated feature engineering work every team does independently. What a feature store is and whether you actually need one.
ML in production: the gap between a pilot and a working system
Why machine learning pilots often fail to become production systems, and what to do differently from the very beginning.
The Transformer architecture: a new universal foundation for sequence processing
What the arrival of the Transformer architecture means for companies thinking about applying language models in their processes.
The gap between an ML experiment and a production system
Why machine learning in a notebook and machine learning in a running product are different tasks with different requirements.