Are You Fed Up with AI Customer Support Guessing Games? Jev and Claude Code Can Fix That

Full Prompt:https://mouen.site/jev-claude-code-pr... 20% Discount Hostinger: https://www.hostinger.com?REFERRALCOD... OpenRouter:https://openrouter.ai/ Is your AI customer support guessing? One confidently wrong reply to an angry customer can cost you that customer. I used Jev and Claude Code to build a customer support app that decides where every ticket goes, and asks a human when it isn't sure. Full build, real numbers. Most AI customer support automations focus on writing replies. This Claude Code tutorial builds the other half: the part that DECIDES — which department, how urgent, is it a refund, does a human need to look. The decisions come from Jev, TypeSafe's small decision model, through OpenRouter. Claude Opus 5.5 only drafts a reply when one is actually needed. What you'll learn: • Stop the guessing: add a confidence gate so unsure tickets go to a human (human in the loop) • Stop overpaying: each Jev decision costs about 200× less than asking Claude Opus • Stop trusting demos: measure accuracy on labeled tickets before you scale • Build it yourself with Claude Code, step by step, and see when NOT to use Jev What I measured in this build: • 82 tickets I labeled myself: 84.1% department accuracy, 22 auto-routed (90.9% of them correct), 60 sent to human review, 2 confidently wrong, 531 ms average, $0.0029 total Jev cost. • 100-ticket batch: 87.0% accuracy, 31 auto-routed (93.5% correct), 69 to review, 0 errors, 49 s for the whole batch, $0.0036 total. • Same 8 examples vs Claude Opus 5.5: both 8/8. Jev ~0.7–1 s and ~$0.00004 per analysis; Claude 4.6 s and ~$0.008 (about 200× more per analysis). The labeled sets are small and hand-made, so treat them as a method, not a benchmark. The thresholds I used (0.80 department, 0.50 urgency, 0.30 human review, 0.40 refund) are starting points to tune on your own data. CHAPTERS 0:00 Cold open: one frustrated customer message 2:00 What Jev actually is 3:41 Choice, Score & Noul 5:05 OpenRouter setup 6:28 Build it with Claude Code 9:52 Live test: an easy billing ticket 11:06 Live test: a technical-support ticket 12:09 The test most tutorials skip 13:32 Add the confidence gate 14:54 Evaluate instead of guessing 18:41 Jev + Claude: the hybrid workflow 21:19 100-ticket batch test 23:42 Jev vs a normal LLM 25:23 When NOT to use Jev 26:32 Final architecture TOOLS • Jev 1.13 by TypeSafe, via OpenRouter (typesafe/jev-1.13-20260917) • Claude Code (Opus 5.5) to write the app • A small local Node.js web app — the "Jev Support Router" LINKS • OpenRouter: https://openrouter.ai • TypeSafe Jev docs: https://docs.typesafe.ai • [Project files / repo, if you publish them] Prices, models and results— check the live pages before relying on them. Music and sound effects are original, made for this video. #Jev #ClaudeCode #CustomerSupport