The author is building a shopping concierge service. Initially it aimed to judge whether vegetables and food are cheap using public statistics, Jev, and LLM. After adversarial review, it shifted toward a mechanism that places scientific promotion criteria between evidence and action.
買い物コンシェルジュを作っていたら、思ったより面白いものになってきた 最初は 「野菜や食品が今安いか？」 を公的統計＋Jev＋LLMで判断する仕組み。 でも敵対的レビューをかけて改修していたら、本質が変わった。 今作っているのは EvidenceとActionの間に“科学的な昇格条件”を置く仕組み に近い
The post showcases a real-time shopping assistant that combines GPT-Live-1 with Jev. As the user speaks, the interface surfaces matching products immediately, even before the turn ends, with Jev using dialogue state to drive the interaction.
Showcases ego lite, Jev, and DeepSeek Flash completing 20 Amazon product decisions in 3.71s versus 54.45s for GPT-5.6 Sol, both scoring 10/10.
Jev scored 3,000 kid snacks on nosugarforkids.com against multiple criteria in 28 seconds for $0.11. Link
