Writing

Notes from production.

What actually happens when LLM systems meet real users, real money, and real data — guardrails, observability, cost, and retrieval.

13 May 2026/RAG5 min read
RAG That Actually Retrieves: Lessons From a £6B Trust Platform
Naive cosine top-k fails on enterprise docs. Here's how chunking, hybrid retrieval, and MCP tool-calling lifted doc retrieval +45% on a platform moving £6B+/year.
04 Mar 2026/LLM Observability5 min read
Self-Hosting LLM Observability When Your Data Can't Leave the Building
For a fintech client, SaaS tracers were a non-starter. Here's the ~500-line custom callback handler into self-hosted Langfuse with cost, latency, and on-prem PII redaction.
22 Jan 2026/Agentic AI5 min read
Stopping a $946 Token Leak with Real-Time Agent Guardrails
An autonomous LangGraph agent was burning ~$946 every two weeks on pathological tool-call loops. Here's the 700-line guardrail layer that killed it.