Lead Data Engineer @ Halter
he/him
Antony has been responsible for Halter's data platform for nearly 5 years and overseen more than 100x growth in customer (and data!) volumes in that time. His proudest achievement to date is leading the team that reduced the AWS component of Halter's cost-to-serve by over 45% in early 2025 -- a significant contribution to the $100M USD Series D raise later that year. Today, Antony's focus remains on getting the most out of Halter's large and uniquely valuable datasets -- particularly in this emergent age of agentic systems.
A year ago I was openly skeptical of the usefulness of AI coding assistants. Oftentimes the result was just frustration; producing queries against tables or columns that don't exist, joins or aggregations that are semantically incoherent, and being confidently incorrect all the while. A surprise then that today I'm one of the heavier users of such tooling at Halter. While the models themselves have improved greatly since my hating days, the biggest change has been my own *mental* model of how to apply them. Over the last 6+ months I've been experimenting and iterating with the approach during my normal data engineering work. In a nutshell, what I learned is that effective use of LLMs in a data engineering context is primarily a *context management* problem. While LLMs have been trained on enough SQL to write syntactically correct queries, they haven't seen the schemas, metrics, or business logic of your particular business -- they don't know what "correct" means _for you_. Deliberate context management is what enables LLMs to discover constraints, correct themselves, and produce results that are actually useful (i.e. correct!). This talk is all about the nuts and bolts of how to actually achieve those outcomes. I'll cover practical examples of teaching LLMs new tricks with "skills", the mental model for reasoning about LLMs (as machines with limited memory), and a few day-to-day phrases you can start using immediately to level up your LLM game. A short live demo will pull all the pieces together and show you the kind of output you can expect when context is managed deliberately in a real data engineering workflow.
