Right now, the common thread is feedback: making it shorter, more informative, and easier to act on. I’m exploring that through retrieval evaluation and interactive explanations of system behavior.
Building
A personal notebook that can hold prose, Python, and small interactive explanations without asking readers to install a runtime. The first example is a convolution kernel explorer.
Thinking about
Evaluation-driven AI systems. How do we tell whether a retrieval change improves the evidence a model receives? How do we make an evaluation set useful without mistaking it for the real world?
The RAG evaluation notebook starts with a deliberately small version of that question.
Experiments
Visual explanations of familiar ideas: kernels, vector similarity, queueing, and tail latency. Building a diagram often reveals which parts of an explanation I do not yet understand.
Other interests
Native tools, Python simulations, and creative coding. Not every useful project has to start with a business case.