One Consultant Built a Tiny AI That Beats GPT at Business Calls

A process-improvement consultant who has worked with Fortune 500 companies got tired of one stubborn problem: in any business workflow, the tricky judgment calls — should we refund this? is this return suspicious? — almost always ended up routed to a human, because writing code that handles messy, ambiguous situations is brutally hard. So he built ImaJev-4b, a small AI that takes text, a form entry, or up to two photos and returns a probability for each possible answer, plus an 'unknown' option, in a single pass.

The first attempt went badly. He trained a bigger model on roughly 500,000 short decisions and it actually got dumber at reasoning, dropping from 64.9 to 42.3 on a hard test, because it had learned to pattern-match instead of think. He then spent two weeks generating genuinely hard questions and keeping only the ones where two different AI models agreed on the answer. That fixed it.

The results surprised him. On JevBench, a niche test for decision-making AI, ImaJev ranked #1 out of 91 entries, scoring 67.37 against 63.29 for the model it was based on. The same week, DecisionBench placed it #3 of 56, ahead of GPT-5.6 Luna and DeepSeek V4.1. He's upfront about the catch: that first-place score weighs accuracy, confidence, speed and cost equally, and on accuracy alone it lands third. Its real edge is honesty — when it claims 90% confidence, it's usually right, and it will say 'unknown' rather than invent an answer.

The whole thing runs on a single Mac or one graphics chip, and cost about $1,200 in rented computing time. It's not a chatbot and it only answers closed, multiple-choice-style questions, so it won't write your emails or plan your trip. But it's a striking demonstration that a cheap, narrowly focused AI can out-perform the giants at one specific job — and that an AI which admits uncertainty is often more useful than a confident one.