Live Agent System

PROBLEM

The paradox of a new AI product: you need proof to win work, and work to build proof. The classic fix is fake case studies and inflated metrics — that collapses on the first real client and burns the brand. On top of that, a technical constraint: zero budget, no payment card, instant replies for every visitor, resilient failover, and abuse protection.

APPROACH

We stopped talking and started showing. On the AI arm’s homepage, visitors describe a task and watch the agent work — steps are honestly labeled, responses come from a real model (proof, not choreography). We built multi-layer failover: if the primary model stalls, the queue falls through automatically — six stages across two independent gateways and three separate model vendors. Quota runs out at the account level, not per model, so the chain switches gateways as well as models — when one shuts entirely, the queue moves to the other. Trust and continuity layers: every conversation is logged anonymously (real demand intelligence), crash alerts email the team, bot protection guards free quota, personal data stays out.

RESULT

The outcome is the system running on ai.ehnuh.com right now: not a mockup, a live model. · Monthly operating cost: ₺0 — built on free model tiers. · Six-stage failover: two independent gateways, three model vendors → assistant stays up if one fails. · The chain is tested regularly with real questions → a free tier vanishing quietly gets noticed. That is exactly what happened in August 2026: all three chain models stopped being free at once, silently. The test caught it; the chain was rebuilt by measurement. · Every visitor question logged anonymously → real demand data. · Bot-protected → automated traffic cannot drain free quota. · Automatic alert on failure → the team knows before visitors do. Lesson: well-built AI speaks through a measurable, running system — not a flashy promise.