We integrated Jev in our chat on public data for Bulgaria (elections, procurements, budgets, politicians etc).

We integrated Jev in our chat on public data for Bulgaria (elections, procurements, budgets, politicians etc). 230+ tool calls, What we measured: - Robustness is where it shines. With typos, Jev alone picks the right tool 94% (EN) / 86% (BG) of the time. Our keyword rules: 28% / 34%. Reworded questions: 86% vs 15–25%. - It cannot fill open values — names, company IDs, free text. It has no primitive that produces one. So it now routes for Gemini 3.5 Flash-Lite: Jev picks the tool, Gemini fills only that tool's parameters. Same day, same 474 questions: parameters right 94.3% / 90.6% vs 83.0% / 77.4% for Gemini alone, with a 4x shorter prompt (3,697 vs 16,150 tokens) and ~0.2 s added median latency. - Weak spot: Bulgarian typed in Latin letters (shliokatitsa) — 71% for Jev vs 94% for Gemini. When Jev is unsure, the question goes to Gemini with the full catalogue. Full results, method and decision history: https://naiasno.bg/chat/evals