How does Backpropagation work?
F(X) = Y describes what a model does once it’s trained. But how does F gets to that state in the first place?
Bay Information Systems
Practical writing on AI and machine learning for people building with it.
Subscribe:.rss.atom.json.md.m3u.pls
Live Last read · last published · next check
Latest episodes
The Four Types of Machine Learning
Machine learning reduces to one expression: F(X) = Y. A model F takes input data X and produces output Y. From these combinations, we can learn all the main strategies to solve data problems with ML.
The Unfinished Notebook
Notebooks are a great way to communicate narrative and encourage engagement in data projects.
Model Choice Determines VLM Inference Cost
Output length bias in the training process impacts the cost of model inference.
Embedding-Based Relationship Discovery: Finding the Real Network in Your Data
Most organisations have a formal model of their customer relationships and a real one that differs from it. Embeddings and community detection surface the real structure.
Engineering in the Exponential Age
Capability is arriving faster than teams can absorb it. The engineering task becomes protocols and interfaces in place of bespoke services.
The Algorithm Lottery
Recommendation systems deliberately surface under-explored content to find new audience clusters. Understanding why this happens changes how you think about content distribution.
How to Structure an ML Team
The four horsemen of the techpocalyse!
What Is NER and Why Does It Matter for PII?
Some tasks are so well understood there are specific classes of models just for that task. Identify named entities and redacting private information are two of those specific tasks.
Falling Token Prices and the Cost of Inference
How do you know if you are saving money when things get cheaper?
Why Private Inference Is Not Fully Private (Yet)
Fully Homomorphic Encryption will bring massive change to the inference landscape, but it is not yet possible.
Multi-Tenant Architecture: A Practical Spectrum
Multi-tenancy is a spectrum from shared tables to fully separate infrastructure, and the right point on that spectrum depends on your risks.
LLM Data Formatting Choices: A Taxonomy
The format you use to pass data to a language model affects reliability and cost more than most expect...
Reheat
Our new open source tool pulls your Search Console queries, embeds them as vectors, clusters by semantic intent, and surfaces ranked content gaps and opportunities as a report.
GPU Market Overview for Local AI/ML Inference: VRAM Capacity and Software Stack Compatibility (Mid-2026)
UK focus with approximate pricing.
Bronze, Silver, Gold: A Layered Data Architecture in PostgreSQL
The medallion architecture maps cleanly onto PostgreSQL. Each layer has a single concern, with transitions being where the interesting engineering lives
A Measurement Framework for Machine Learning Projects
A practical framework for defining, building, and maintaining an ML system -- including a free checklist!!
Event-Driven Embeddings in PostgreSQL: async and LISTEN/NOTIFY
Postgres LISTEN/NOTIFY mechanism can trigger embedding generation on INSERT, without polling, without a separate scheduler, and without coupling services.
Don’t Build Models, Build Measurements
If an organisation can define "what correct looks like" ML projects will likely succeed; it also answers the gnarly "Are you automating my job?" question
What Is Data Engineering?
Two Jobs, One Title; which one do you love most?
Open Weight Models and the Separation of Concerns in AI
Often framed as an ideological discussion, open weight models actually solve the business problem which ruins closed model providers over the long run.