AI / ML
RAG Retrieval Optimization: Precision, Not Recall@K
An LLM's answer is set by the precision and order of its context, not by Recall@K. Retrieving more documents, badly ordered, makes RAG answers worse.
Writing
On machine learning, LLMs, and building applied AI systems, plus notes on the craft of engineering.
AI & Machine Learning
AI / ML
Reduce LLM API Costs: Input Tokens Dominate the Bill
An LLM bill is a product of a few factors, and input tokens re-read on every call dominate. The biggest wins are lossless discounts, not shaved output.
AI / ML
AI Agent Design Patterns: Why Self-Critique Loops Fail
Self-correction degrades GPT-4 on GSM8K from 95.5 to 89.0. The agent patterns that work all share one thing: an oracle the model cannot fake.
AI / ML
AI Data Pipelines: Why Embeddings Break Your Audit Trail
ETL fails loudly; AI pipelines fail silently. Chunking and embedding destroy what you can audit, so contracts, quality gates and lineage must run upstream.
AI / ML
8 AI Model Types Explained: LLM, MoE, VLM, SLM and More
What LLM, MLM, SLM, LCM, MoE, VLM, SAM and LAM actually mean, how they differ, and why the eight names collapse into four design decisions.
AI / ML
Building AI on AWS in 2026: The Four Decisions That Matter
AWS now lists about ninety AI services. Four decisions determine the rest: vector store, agent runtime, model binding and the silicon underneath.