Category: Data Science
-
MLflow Experiment Tracking, Model Registry and Versioning (Data Science Series, Part 24)
A model you cannot reproduce is a rumour. Here is how I set up MLflow tracking, a model registry and aliases on the churn project, including the errors you will hit on a local file store.
-
Feature Stores and Training Serving Skew in Machine Learning (Data Science Series, Part 23)
A model that scores well offline and badly in production is usually not a modelling failure. It is two pieces of code computing the same feature differently. Here is how skew happens, what a feature store fixes, and when a shared library is the better answer.
-
Serving Machine Learning Models: Batch, Real Time and Streaming (Data Science Series, Part 22)
Batch, real time and streaming are three different answers to the same question, and picking the wrong one costs money for months before anyone notices. Here is how I choose, with runnable FastAPI and batch scoring code for the churn project.
-
From Notebook to Python Package: Structuring a Data Science Project Someone Else Can Run (Data Science Series, Part 21)
A working churn notebook is not a deliverable. Here is the src layout, pyproject.toml, path handling, test suite and artifact versioning that turn 175 lines of model code into something a colleague installs and runs in one command.
-
Recommender Systems in Python: Collaborative Filtering and Cold Start (Data Science Series, Part 20)
Collaborative filtering in Python with implicit, evaluated honestly by leave one out ranking. A popularity baseline beat matrix factorisation, and a new user returned five items scored exactly zero.
-
Time Series Forecasting in Python, and Why Your Train Test Split Is Wrong (Data Science Series, Part 19)
A shuffled cross validation split on time ordered data will hand you a score you cannot ship. Here is what an honest forecast split looks like, how to backtest with a moving origin, and how far ahead a model is actually good for.
-
NLP in Python, From Bag of Words to Transformers (Data Science Series, Part 18)
Four ways to turn text into features, measured on the same split: counts, TF-IDF, frozen embeddings and a fine tuned transformer. With the latency and cost each one buys you, and where text features belong in a churn model.
-
PyTorch Essentials in Python: Tensors, Autograd and a Training Loop You Can Read (Data Science Series, Part 17)
Tensors, autograd and a hand written training loop in PyTorch 2.13, rebuilt on the churn model from Part 16. Includes the three errors that cost me the most time and an honest CPU versus GPU comparison.
-
Neural Network Fundamentals in Python, Built Up From Logistic Regression (Data Science Series, Part 16)
A neural network is logistic regression with a layer in the middle. I build one in thirty lines of numpy, then show the scaling mistake that dropped average precision from 0.5445 to 0.1062 on the churn model.
-
Hyperparameter Tuning in Python: Halving Search, Optuna and Honest Model Selection (Data Science Series, Part 15)
A random search, a halving search and an Optuna study run against the same churn model, with wall times, the honest test score, and the nested cross validation number that showed how much of the gain was imaginary.
-
Imbalanced Data and Resampling in Python Without Fooling Yourself (Data Science Series, Part 14)
Rare positive classes break accuracy, flatter ROC AUC and invite resampling mistakes that leak. Here is what actually moved the money on a churn model, with real numbers for SMOTE, class weights and threshold tuning.
-
Clustering and PCA in Python: Unsupervised Learning on Real Data (Data Science Series, Part 13)
K means and principal component analysis on the same 486 row churn frame we have modelled since Part 6. Real silhouette scores, four profiled segments, and an honest test of whether cluster labels help a classifier at all.
Architect’s Toolkit
PJ’s Tools
VMware Cloud Foundation
- VCF Documentation
- VCF 9 Planning & Preparation Workbook
- VCF Bill of Materials (BoM)
- VMware Compatibility Guide
- VMware Interoperability Matrix
- VMware Configuration Maximums
- VMware Ports & Protocols
- VMware Hands-on Labs
- RVTools Download
Nutanix
AI & Cloud-Native Platform
- NVIDIA Build (Model Catalog)
- NVIDIA AI Enterprise Reference Architecture
- NVIDIA NIM Performance Benchmarking
- NVIDIA NGC Catalog
- NeMo Microservices Helm Chart
- Helm Charts Repository
- Hugging Face Models
Architecture & Design
About the Author
Dr Pranay Jha
Dr. Pranay Jha is a Cloud and AI Consultant with 18+ years of experience in hybrid cloud, virtualization, and enterprise infrastructure transformation. He specializes in VMware technologies, multi-cloud strategy, and Generative AI solutions. He holds a PhD in Computer Applications with research focused on Cloud and AI, has published multiple research papers, and has been a VMware vExpert since 2016 and a VMUG Community Leader.
You May Have Missed
DrJha