Blog
Notes from the build.
Agentic RAG and GraphRAG: when basic retrieval is not enough
Single-shot vector search fails on multi-hop, whole-corpus and vague questions. What agentic retrieval and GraphRAG each fix, and what they cost to run.How to run an AI proof of concept that survives real data
How to run an AI proof of concept: test the riskiest assumption on real data, set pass and fail first, with a week-by-week plan and a one-page brief.Cloud management for growing teams: governance, cost, uptime
What to put in place between one cloud account and a platform team: accounts, identity, infrastructure as code, on-call, backups, cost and a staged checklist.Build vs buy for AI: a decision framework
A framework for AI build vs buy decisions with three options: buy a product, assemble from platforms and APIs, or build custom. Includes a decision table.Vector search explained: embeddings, indexes, distance metrics
How vector search works under the surface: embeddings, cosine versus dot product, HNSW and IVF indexes, quantisation, filtering and how to measure recall.AI-ready data: what to fix before you build anything
A practical audit of the data behind an AI project: owners, permissions, freshness, duplicates, scans, personal data and test cases, with a scored checklist.Predictive maintenance with IoT and AI: a practical guide
How to build predictive maintenance that planners trust: which signals to measure, how to cope with few failures, and how to prove value on one asset class.Prompt engineering: what still matters and what does not
A practitioner's view of prompt engineering now: the habits that still improve results, the tricks to drop, a before and after prompt, and how to test.Retrieval-augmented generation (RAG): how to build it well
How retrieval-augmented generation works stage by stage, the choices that decide answer quality, how to diagnose failures and what a first version needs.