An experiment in creating a fine-tuned SLM/LLM model as a proof-of-concept using Apple's MLX framework. Required ingredient: MacBook with Apple silicon that
supports unified memory and Ollama

"Our dbt DAGs were spending 81% of their time on parsing overhead. By switching to consolidated dbt build commands with tag-based selection, upgrading Python, and right-sizing our warehouse, we saved 37% on compute costs annually."

A case study describing the business rationale, technical challenges, and impact of our migration to Node type stripping

One approach to rigorous Bayesian power analysis in a difference-in-differences related quasi-experiment.

As engineers, we like boolean logic. We like yes or no answers. We like right
answers. So, when coming across problems that do not have right answers, it
invokes a different pattern of thinking that we don’t use every day. One of
these such problems is deciding whether you should spend the time engineering
a task, or just do it manually.  I think as engineers we often default to the
technical solution, even if it means spending way more time putting together
the pieces to automate something vs just copy pasting something 50 times.

How to be productive in a totally novel domain with little knowledge, using Artificial Intelligence.

A cautionary tale of looking for a technical solution for Node unit test performance optimization where an easier less technical solution exists.

User generated playlists has been the number one requested feature since I joined Calm 2.5 years ago. However, during that same time span, it was never prioritized high enough to make it onto a product roadmap. So, like a naive kid new to the corporate world, I set out on a journey to make playlists happen. How hard could it be?