AI & ML Engineering
Building and shipping systems whose behaviour is probabilistic: LLM application architecture, retrieval, evaluation, agents, inference economics and the operational discipline underneath.
Ship a non-deterministic system you can still be accountable for.
Foundational
3The Demo To Production Gap
A convincing prototype is a small fraction of the work. What sits between a demo that worked and a system you can operate, and how to reset executive expectations after they have seen the demo.
9 min readexpectationsproduction readinessThe Six Layers Of AI
Classical AI, machine learning, neural networks, deep learning, generative AI, agentic AI. The six names are the right six, and the popular graphic gets one thing wrong that changes what you conclude from it. What each layer is, what it cannot do, and which to reach for.
7 min readfundamentalsmachine learningThe Three Jobs Of Labelling
One ordinary word doing three unrelated jobs — measuring, teaching and operating — with opposite requirements. The question that separates them, why mixing measurement data into training flatters rather than breaks, and how to read a number somebody else produced.
8 min readlabellingevaluation
Practitioner
4Evaluation Harnesses For LLM Systems
The piece almost every team skips. How to build a golden set, design a rubric, use LLM-as-judge without being fooled by it, and turn "is this better?" from an argument into a measurement.
9 min readevaluationgolden datasetsRAG Architecture That Survives Production
Retrieval quality sets the ceiling on answer quality, and chunking decides retrieval quality. Hybrid search, reranking, grounding and citation, re-indexing, and how to tell which half of the system broke.
10 min readragretrievalTesting Non-Deterministic Systems
Assertion-based testing assumes one correct output. Model-backed systems do not have one. What replaces equality checks, where deterministic tests still belong, and what the test pyramid becomes.
9 min readtestingnon-determinismWorld Models And The Convergence Claim
Two ways to build a predictive system — predict the next token in the data, or predict in representation space. A recent theorem says encoders trained on different modalities converge on one embedding, and the coverage has outrun what was proved. What the result licenses, what it excludes, and how to read a claim like this when the next one arrives.
6 min readworld modelsjepa
Other domains