Case studies
Each of these is one system I built, written as an evidence pack: the problem, the architecture, the result, and what it proves. They are the proof behind the method. Numbers come from the code, not a brochure.
-
SignalTrace
SignalTrace is thirteen independent research wikis published from Obsidian vaults through a custom multi-wiki build on Quartz: 912 markdown files, 780+ entity pages each carrying an evidence grade. It is the method published.
Read the case study -
JobAbroad
JobAbroad is a paid, members-only portal for South Africans planning to work overseas. It wraps a research vault in semantic search with cited answers, a readiness assessment, and an AI coach, built on Next.js and Supabase with pgvector.
Read the case study -
EdenFintech scanner
A Python equity-research scanner that puts an LLM analyst panel behind a type-enforced information barrier, grades its own evidence, logs every model call, and scores stocks with plain arithmetic. The same discipline as my research-graph method, applied to quant.
Read the case study -
WeCoza 3.0
WeCoza 3.0 is a full line-of-business app built as a WordPress plugin for a SETA-regulated adult-education provider, with its operational data and rules living in a 44-table PostgreSQL schema reached through a hand-rolled PDO layer that bypasses WordPress entirely.
Read the case study -
Skills authorship
Authoring a skill means writing the rules, scripts, and guardrails an AI agent reads before it acts, so it follows a project's real conventions instead of being re-prompted each time. I have authored eight personal workflow skills, three larger engines (vault-builder, infranodus, graphify), and a full Claude configuration embedded inside a client codebase. This is the harness-over-prompt thesis with the artifacts behind it.
Read the case study
