Where AI in UX actually saves me time
Two years of daily use, boiled down: AI in UX is good at sorting research transcripts, flagging the obvious usability problems and drafting copy from client bullets. The judgment stays with you.
Tag: AI Assistants
RAG chunk size is the variable most people never tune
Small chunks lose context, medium chunks produce near identical similarity scores, large chunks are stable but noisy. How RAG chunk size affects retrieval, plus a PHP splitter that respects word boundaries.
Reinforcement learning for LLM reasoning: proofs over vibes
Part 2 of the vibe proving series: how an RL loop built around a hard proof checker teaches a model to follow logic rules, and where it still falls over on nested proofs.
Why I use GliNER2 for structured data extraction
A client had ten thousand messy biographical records that needed to become a clean knowledge graph. GPT-4o would have done it and cost a fortune, so I used GliNER2 instead: schema-driven extraction that runs on a CPU.
How I handle covariance shift with IPW
A WooCommerce recommendation engine scored 0.85 AUC in staging and fell apart on launch day because the live audience did not match the training data. Inverse probability weighting is how I got an honest read on what the model was really doing.
Specialized data roles beat the generalist stack
Data work is moving away from generalists and toward people who know the domain, whether that is archaeology, renewable energy or sport. I got this wrong once on a solar reporting engine, and the fix was domain logic rather than a better query.
A tabular foundation model beats XGBoost on 800 rows
Predicting lead conversion from 800 WooCommerce rows, where XGBoost just memorized the noise. Why TabPFN and in-context learning made a better baseline than another afternoon of hyperparameter tuning.
Why I swapped mega-prompts for a custom AI assistant
I spent hours refining a mega-prompt and still got different output every session. Here is why that logic moved into a custom AI assistant, how the MATCH framework structures one, and when a private local session is the better call.
The YOLOv1 loss function, one term at a time
A client’s detector found the gearboxes and ignored the bolts, and the problem was the loss function rather than the dataset. Here is how the YOLOv1 loss splits into five terms, why width and height get a square root, and what the lambdas are for.
Why MRR lied to me about a WooCommerce search
MRR said a WooCommerce store search was fine while shoppers kept skipping the top result. Here is why binary relevance metrics mislead, and how NDCG and ERR score a result list closer to the way people read one.