I spent this week trying to break my own AI system. mostly succeeded lol. warrant is an ops assistant I built for a fictional hospitality studio — answers questions, looks up clients, drafts emails. nothing that reaches a real customer sends without a human approving it first. sounds simple. getting it to feel trustworthy wasn't. first version: asked it "did we send that email today?" — it tried to draft a brand new email instead of just answering the question. asked it retainer pricing — it gave me a vague non-answer instead of the actual number sitting right in front of it. approved an email — it told me "sent successfully" before actually checking if it had been. none of those are AI problems. they're architecture problems. so I rebuilt the decision layer: the system now has to ask itself "is this a new action, or a question about something that already happened" before it's allowed to touch a single tool. it verifies an action actually worked before reporting success. and when I asked it something it genuinely can't do (list every customer in the CRM) it didn't fake an answer. it said so, and told me what it could do instead. picture a small studio running this in-house: every client email drafted and logged automatically, nothing sent without a human glance, and a full trace of every decision the system made, watchable live, not buried in a log file you check after something's already gone wrong. that's not a hypothetical productivity number — that's an afternoon of admin work turned into two minutes of approvals. this is the part most AI demos skip: showing you when it was wrong, and what it took to make it trustworthy instead of just impressive. built with groq, supabase, and pgvector. live: usewarrant.vercel.app