I use these projects to test specific claims about inference and agent reliability. Some became working prototypes. Others produced a useful negative result and stopped there.
The write-ups focus on what I was trying to learn, what the evidence supports, and where the boundary of each project sits today.
Experimental: local-first inference governed by explicit policy, measured capability, and strict tool authority.
Proof of concept: human-approved incident remediation with bounded authority and an observable execution trail.
Prototype: keeps coding-agent dispatch decisions in versioned repository state instead of chat history.
Pre-release: measures deployed model behavior with deterministic grading instead of treating a model name as a benchmark.
Experimental: treats structured output and tool use as validation boundaries, not prompting conventions.
Experimental: makes the ReAct loop typed, bounded, and observable, including how it stops.
Concluded experiment: follow-up tests showed the apparent tool-call problem was mostly a serving-stack artifact.
Concluded prototype: worked on a narrow raw-output path; later evidence showed modern parsed endpoints usually do not need it.