rebecca writings thoughts prayers talks ships

writings

What shipping a production agent actually looks like

In high-stakes AI, self-correction is the wrong default

Building Faro

'Be careful with tool output' isn't a security strategy

One slow MCP server degrades your whole agent

Multi-agent is a cost, not an upgrade

A small reranker lift on a specialized corpus is a diagnosis

The RAG decision you can't undo

Citations don't mean grounding

LLM-as-judge has four failure modes

Bad RAG answers have three separate root causes

Retrieval quality is the ceiling. The model is not.

LLMs from Scratch - Pt. 5

LLMs from Scratch - Pt. 4

LLMs from Scratch - Pt. 3

LLMs from Scratch - Pt. 2

How I built an LLM from scratch

My Journey Into Agentic AI