Real production bugs, fixes, and lessons from building VectorMatch. First-hand engineering war stories.
Five undocumented traps in Shopify’s Global Catalog API that make every AI visibility score a lie. 30 rounds of testing, 98.7% recall, and the invariants that caught my own tool hallucinating data.
Seven months of silence. Eleven applications. Four interviews. One contract. A field report on bypassing ATS latency, the multi-discipline bypass, and what five AI agents found that I couldn’t.
Everyone reporting AI visibility numbers is reporting a sample statistic as a fact. I did this job by hand for Google in 2008, and the discipline was better then.
The LLM is the most expensive stage in a matching pipeline, not the interface to it. I built a 3-gate funnel where 95%+ of jobs are filtered by SQL and vector similarity before any model call. Here are the real numbers.
Six parallel Inngest workers each held their own rate limit counter, blowing through a 2 req/s cap. Tests passed because they were sequential. Redis fixed it.