Notes on applied AI
Field notes on machine learning, generative AI, and the engineering it takes to put models into production — written for the people who have to trust them.
-
When the generator outgrows the verifier
A follow-up to June's GraphRAG + Chain-of-Verification post. I swapped the 70B generator for a frontier model (Claude Fable 5) and re-ran the whole pipeline — and the verification layer, not the generator, became the weakest link. This post covers what the frontier model fixed, the false-positive failure mode it exposed in CoV, and a "verify-and-repair" redesign that fixes it.
Read -
Supplementing RAG to improve causal analysis
In some cases, particularly with topics spanning multiple contexts, current LLMs often struggle with complex cause-and-effect reasoning. In this post, I walk through an example of that failure mode, an implementation of GraphRAG (using Neo4j), and a framework for evaluating whether specific causal claims are actually supported by evidence.
Read -
Evaluating RAG faithfulness without fooling yourself
A single accuracy number hides more than it reveals. Here's the metric we actually trust for retrieval-augmented generation — and the math behind it.
Read