AI experiments
Beyond the Model
I build small experiments to understand how AI systems behave when they have to produce useful, checkable results. These write-ups cover coding agents, local screenshot search, and comparisons with models designed for other kinds of data.
Experiments and results
Coding agentsChecking when a coding agent is doneTwenty local-model runs led me to review the harness, add a completion check, and repair the evaluator. Results and run records are included.Local AICan AI help me find information in my saved screenshots?Testing local screenshot search with fictional messages, bookings, and receipts. The write-up separates retrieval from answer accuracy; phone performance remains untested.Model comparisonsLooking beyond LLMsLocal case studies comparing a language model with classical ML and specialized models across different kinds of data. Methods, results, and limitations sit alongside the experiments.
Reading the results
These are small studies with specific tasks, datasets, and runtime limits. Each write-up explains what was tested, what went wrong, and what remains uncertain. The records let you inspect the work behind the conclusions.