~/writing
Writing
Notes and teardowns on AI-native business intelligence - how AI is reshaping data analytics and BI for the teams that run on it.
Adding an Evaluation Gate Inside a Fabric Notebook
Fabric can handle the OpenAI authentication for us. A small experiment using GPT-5.1 to review a data agent’s answer, with code checks deciding what can move forward.
Text-to-SQL Is Harder Than It Looks
Text-to-SQL demos beautifully and breaks quietly. The model nails the syntax — then trips on ambiguity, your schema, and answers that look right but aren't. A field guide to the gap between the demo and a number you'd actually trust.
The Window Is Not the Memory
Retrieval was fixed — then the chatbot got dumber the longer a technician talked to it, and forgot everything by morning. The two things people conflate: the context window and memory.
What Actually Breaks Agentic RAG in Production
I built an agentic RAG system for a cleaning-robot company's field technicians. The agent was the easy part — here's what actually broke in production, and the four fixes that mattered.