Notes on human–AI collaboration, tools for thought, and doing more with less.
A small always-on server, dark to the internet, that keeps an Obsidian vault converging and version-controlled, hosts coding agent sessions I can drive from a phone, and — unplanned — reaches my tasks, mail, and calendar.
AI-assisted coding accelerates one pillar of development work — writing code. The real leverage is an assistant that works across all three pillars, project management and infrastructure included, closing the whole loop instead of one third of it.
Integrating decision support systems into business processes — the building blocks of descriptive, predictive, and prescriptive systems, and why results happen in processes, not in the systems themselves.
A conceptual map of artificial intelligence — from the definition of intelligence to the five abstraction layers of every machine learning solution.
Linear regression with several features in matrix form: the closed-form least-squares estimator derived and computed, the sampling distribution of the coefficients, and a from-scratch look at LASSO, ridge, and elastic-net regularization.
Fitting a line from scratch: the model and its assumptions, least-squares estimates derived by hand and by optimizer, the sampling distributions of the coefficients, a confidence band for the mean response, and R².
Three estimators for the mean of a normal sample — sample mean, least squares, and maximum likelihood — derived from scratch and shown to coincide, then confidence intervals checked against the truth.
Why the average of many trials settles on the expected value — shown with coin and dice simulations in NumPy, then the weak law stated and proved via Chebyshev's inequality.
A from-scratch look at Markov chains with NumPy — the Markov property, transition matrices, and how any starting distribution converges to the same stationary distribution.