Notes from moving a Connect Four agent off tabular Q-learning and onto a DQN: why correlated updates destabilize the network, how action masking handles full columns, where the GIL caps throughput, and why the agent learned attack before defense.
An LLM can retrieve the right document and still quote the wrong price. This is the detection and healing layer I built to catch that: five failure patterns, a faithfulness score, deterministic patches, and a SQLite drift monitor that caught a real bug in staging.
LangChain is fine for a prototype. Once there is real traffic, several agents coordinating and a pager attached, the abstraction costs more than it saves. This is the case for writing the orchestration layer yourself.
Ridge, Lasso and ElasticNet land within 0.3% of each other on prediction accuracy, but Ridge fits eight times faster and Lasso’s feature selection falls apart once features are correlated. Three numbers tell you which one to fit.
Hard-coded constants assume a perfect environment. This covers four ways to model uncertainty, from robust optimization to two-stage recourse, and refactors a fixed cron batch size into one that reacts to load.
High accuracy is the easiest number to fake. This post covers the machine learning pitfalls I keep finding in production systems: library defaults nobody checked, data leakage from scaling before the split, and metrics that ignore what the model actually decides.
The data science career path in 2026 asks for more than Python. Notebook work is the easy half; the rest is MLOps, domain knowledge, and keeping your own judgment while agents handle the tedious parts.
AI has absorbed the task-level work, and juniors who sell themselves as code writers are getting ghosted. After 14 years in WordPress development, here is what I look for instead: work you actually owned, and a repo that shows a human debugged it.
Most WordPress resilience testing is refreshing the checkout page and hoping. Intent-based chaos engineering starts from a stated hypothesis about behavior instead, with fallback paths around external calls and abort signals tied to checkout completion rather than p99 latency.
Ghost is an agent-first Postgres platform. Notes from using its MCP server and database forking to let AI agents test schema changes and index strategies without touching a shared staging database.