Merchant Ops Dashboard + Voice Agent

The problem

AI voice agents are hard to evaluate. Unlike text, there is no cheap way to tell whether a change to prompt or model made the agent better or quietly worse across the calls that matter.

The approach

Built a merchant management dashboard alongside a testing framework for AI voice agents. Next.js and FastAPI serve merchant operations visibility, while the evaluation harness drives TTS and STT through scripted scenarios to score agent behaviour automatically.

Results

• Operations visibility for merchant management in one dashboard.
• Automated voice-agent evaluation rather than manual spot-checking.

Engineering notes

Treating the voice agent as an unreliable dependency — scored against fixed scenarios on every change — is what makes iteration safe.