Experimenting with AI before recommending it

How do you give a defensible answer to “can we use a small local model for this?” — when existing evaluation tools are built for engineers shipping production systems, not strategists making feasibility calls?

The hard part of AI feasibility work is not running the model. It is structuring the noticing — and that is a design problem, not an engineering one.

A local-first evaluation environment running on PHP, SQLite, and vanilla JS — talking to local models through Ollama. Projects, test cases, test inputs, and test rounds, with human-in-the-loop scoring on quality and accuracy.

In active use across four real test projects and five small models. Produces defensible feasibility recommendations at the pre-commitment stage — before any infrastructure is written.

How can AI integrate directly into compliance workflows without replacing human decision authority?

Early-stage AI systems require clear decision boundaries - assistance accelerates work, but authority must remain human.

Working product with structured entity logic, AI document processing, and compliance workflow modeling. AI integrated through automated summarization, context-aware overviews, and structured tagging.

Shifted compliance from document storage to structured, AI-supported intelligence. Marks my transition from interface-focused UX toward AI workflow and product architecture thinking.