writing
I write about the engineering and product decisions that make AI systems useful in production: retrieval, evaluation, agent design, reliability, and user experience. New essays land on One Token at a Time ↗.
The model matters, but so do the first message, reply length, ambiguity handling, and whether the flow is controlled by code or left to the model. Notes from building a bilingual screening agent—and from giving the LLM one job: phrasing.
companion project Lucia, the system behind the essay →
Technical writing published during my time at Neodata: where retrieval-augmented generation creates business value, and how prompting practices such as few-shot examples and structured reasoning shape system behaviour.