Ongoing experiments

A short ledger of the experiments that shaped this portfolio, including the ones that didn't work and the ones that changed direction halfway through. The aim is to make fail fast and cheap a visible practice rather than a slogan.

#	Experiment	What I tried	What I found	Status
1	Writing for AI	Restructured the first NimbusWiz pages with explicit chunk markers and metadata	Chunk markers were the right idea. The specific syntax went through three iterations before settling	Restructured all NimbusWiz pages; built the evaluation dataset; wrote the chunking playbook
2	RAG chatbot	Built the full RAG stack against the restructured docs	The architecture works on paper. Running the eval requires paid API access, so the causal claim about restructuring vs retrieval lift is still pending	Chatbot architecture ready; the comparative measurement is still open
3	AI feature doc plan	Drafted a doc plan coordinating ML Engineering, Trust and Safety, Product, and Marketing for a hypothetical AI feature	The coordination work is the documentation work. The artifacts are outputs; the conversations are the substance	Still open: adding the plan as a template

Each row links to the full case study.