Many Small Models for Speed

Many Small Models for Speed

 Posted on November 10, 2024  |  3 minutes  |  496 words  | 

LLMs are pretty cool, but they are massive! If you want to run those for yourself you need a hefty GPU and quite a lot of engineering. But the world of machine learning is so much bigger then LLMs. In organizations all over the world, there are models forecasting time-series, predicting prices, creating embeddings, classifying categories and what not. If you have several prediction/classification steps that combine into one end- result, you could consider training one bigger model that does all of the things. [Read More]
MLOps  deployment  production 

A Model not in Production is a Waste of Money and Time

A Model not in Production is a Waste of Money and Time

 Posted on November 2, 2024   (Last modified on October 31, 2024)  |  2 minutes  |  401 words  | 

I always push on people to make their ML project reach production. Even if it is not that good yet and even if you could eke out a bit more performance. I’ve been inspired by the dev-ops and lean movements and I hope you will be too. ML products have many ways to improve, you can always tweak more. But ML is high risk, with a possible high reward and relatively expensive compared to ‘normal code’. [Read More]
MLOps  production 

Your Machine Learning Model is not the Product

Your Machine Learning Model is not the Product

 Posted on November 1, 2024   (Last modified on October 31, 2024)  |  2 minutes  |  253 words  | 

I’m so sorry. Your precious AI model, with handcrafted beautiful perfect features, with awesome hyper parameters, is not the product. Listen, it is awesome work, not a lot of people can do it, but a good ML model is not the end-product1. I want to talk about value. In the jobs I’ve worked the machine learning model was part of a larger system. And only when all the components come together you create value. [Read More]
MLOps  Value-Stream 

Just enough kubernetes to be dangerous

Just enough kubernetes to be dangerous

 Posted on July 30, 2024  |  2 minutes  |  401 words  | 

As a data scientist that wants to achieve production results, one of the best options is to make your work available in kubernetes. Because kubernetes runs on all clouds and because many organizations use kubernetes. Make your prediction API available in kubernetes and your organization can ‘just’ plug it into their systems. Many data scientists don’t know anything about docker, not to mention kubernetes and its main tool helm. I think you should learn and practice just enough helm to be dangerous1. [Read More]
data_science  kubernetes  MLOps  production 

High and Low Variance in Data Science Work

Consistency or peaks, pick one

High and Low Variance in Data Science Work

 Posted on March 6, 2023   (Last modified on August 13, 2023)  |  2 minutes  |  248 words  | 

I recently read “High Variance Management” by Sebas Bensu and this made me think about datascience work. First some examples from the post: Some work needs to be consistent, not extraordinary but always very nearly the same. Theatre actors performing multiple shows per week need to deliver their acting in the same way every day. Their work is low variance. Some work needs superb results, results you don’t know if you can reach it but you try it many times and between all of the failures, you might find gold. [Read More]
MLOps 

Are you a Fearless Deployer?

Fast experimentation and confident deployments should be your goal

Are you a Fearless Deployer?

 Posted on November 21, 2022  |  2 minutes  |  422 words  | 

how do you feel when you press the ‘deploy to production’ button? Confident, slightly afraid? I bet many data scientists find it a bit scary. It’s worth it to dig a bit deeper into this fear. In my ideal world we are not scared at all. We have a devops mindset. We have no anxiety, no fears at all. You should be confident that the deployment pipeline takes care of everything. [Read More]
deployment  iteration  MLOps  training-pipeline 

Do you Need a Feature Store?

From simple to advanced

 Posted on October 22, 2022   (Last modified on October 23, 2022)  |  5 minutes  |  1025 words  | 

A feature store is a central place where you get your (transformed) training and prediction data from. But do you need this? Why would you invest (engineering effort) in a feature store?1 All engineering is making trade offs, a feature store is an abstraction that can lead to more consistency between teams and between projects. A feature store is not useful for a single data scientist for a single project. It becomes useful when you do multiple projects, with multiple teams. [Read More]
mlops  featurestore  intermediate  tools 

Reading in your training data

Data Ingestion Patterns for ML

 Posted on August 15, 2022  |  8 minutes  |  1493 words  | 

How do you get your training data into your model? Most tutorials and kaggle notebooks begin with reading of csv data, but in this post I hope I can convince you to do better. I think you should spend as much time as possible in the feature engineering and modelling part of your ML project, and as little time as possible on getting the data from somewhere to your machine. [Read More]
development  explainer  mlops  quickthoughts  advanced 

Data Science Technical Terms: Job Titles and Fields

MLE, AE, DE, DS, WTF?

Data Science Technical Terms: Job Titles and Fields

 Posted on January 29, 2022  |  3 minutes  |  556 words  | 

What do I mean when I talk about MLops, Machine Learning Engineering, or data science? I call myself data engineer, data scientist, or machine learning engineer. But never an analyst. To me these job-titles all have a certain meaning, although they overlap. Here is what the job titles mean to me, right now. The first thing you need to keep in mind is the size of the organization the size of the data team and the data-maturity of an organization. [Read More]
data_science  mlops  jobtitles