Things I figured out while building. Mostly on-device AI, VLM evals, prompt iteration, and the occasional side-project postmortem.
A ~600-line Swift tool that proposes (prompt, decode-params) cells, scores each on a labeled set, and emits a Pareto front. So I stop hand-rolling v6.
Four MLX-quantized VLMs, four prompts, 56 hand-labeled scans. What moved the number, what didn't, and one model that was completely dead across every prompt I tried.
Shift focus from what you want to communicate to what your audience needs to hear.
Personal study notes on Google Cloud Platform fundamentals, published gradually.
A 16-week guide spanning GCP foundation, data science, backend, frontend, DevOps, and security.
Step-by-step setup for Python Poetry with Visual Studio Code on Windows and Mac.
Technical deep dive into modern CTR prediction systems and the math underneath.
A walk-through of the DCN v3 paper and what it changes for click-through-rate prediction.
Cascading recommendation systems for efficient large-scale ad recommendations.
First in a management series: building cultures of ownership and results.
Adapting Agile story-point estimation to data science teams and projects.
On Hernando de Soto's theory that legalizing informal property rights unlocks economic development.