Welcome 🌟

An introduction to this website and a brief personal profile.

Notes from after I entered credentials on a convincing fake government site: what I did to contain the damage, file reports, flag my credit files, and keep a paper trail.

Problem set on Gaussian quadratic forms, OLS vs MLE, BLR posteriors, design matrices, weight-space GPs, predictive covariances, and the BLR–GP equivalence.

Problem-style sheet on VI evidence lower bounds, ELBO–KL decompositions, mean-field updates, EM for a two-component Gaussian mixture, and ELBO-based views of EM.

Problem sets on ELBO vs KL to the true posterior, mean-field restriction, coordinate ascent updates, a small graphical model, EM vs VI, and diagonal Gaussian encoders in VAEs.

The more I use AI, the less I believe the future belongs to isolated prompts or isolated models. A month ago, it still felt possible to keep up with the frontier by watching new releases, trying new workflows, and refining protocols by hand. I even designed AGN around that assumption: a coordinated multi-agent structure with differentiated roles, explicit orchestration, and a clear human center. But the pace has changed. New agent platforms, new integrations, new toolchains, and new workflow abstractions now appear almost daily. Some ideas that felt rare not long ago are already showing up elsewhere in convergent form.

A compact final-exam review of the definitions, interpretation language, and modelling ideas that must be known even when the formula sheet is provided.

A full worked exercise: joint and complete-data log-likelihood, responsibilities, ELBO decomposition, M-step updates, and why EM is a special case of VI.

Common AI-assisted LaTeX habits that break display math on Hugo sites: invalid =, Setext headings from lone equals signs, and minus-on-its-own-line—plus concrete fixes.

How splitting roles across agents—architect, reviewer, and scout—reduced cognitive load while restructuring a project during exam prep.