Epsilon AI Console

A support manager who is not a developer has to run an AI agent safely: see how it answers, fix what it gets wrong, check a fix before customers see it and undo it if needed. Inside the help desk the bot must answer every customer message exactly once, hand over to a person cleanly and follow channel rules such as Meta’s human-agent tag. A design review showed that the first logic could leave a message unanswered or answer it twice.

The Next.js 16 app has five areas: Monitoring (dashboard, request log, unanswered questions), Studio, Evaluate (a tester, A/B comparison of two configurations, saved test sets), AI settings and Integrations. In the Studio an editing agent built on the Claude Agent SDK proposes changes to the knowledge base and rules, and each change needs approval, is versioned and can be undone. The interface is in Georgian, English and Ukrainian, light and dark, on shadcn/ui with WCAG 2.2 AA gates and a copy glossary; before that rebuild, four reviewer agents ran a read-only UX audit with 84 findings (12 P1, 47 P2, 25 P3), fixed in phases.

The agent runs on Anthropic or Google models with a backup model, and its rules and knowledge base are cached in the prompt. A conversation engine classifies each message, holds a turn lease so the message is answered once, retries, checks delivery and keeps to deadlines. Bot mode hands chats to selected online operators through a queue, respects working hours, gives the operator a summary in their own language, closes finished chats and reads photos and files. It all runs in Docker behind Caddy on the client’s server, with daily backups.