Sellnseal Voice AI
A voice agent that replaces the manual call-center step e-commerce sellers use to confirm cash-on-delivery orders before shipping.
Role
Solo build — architecture, voice pipeline, and telephony integration.
Context
Cash-on-delivery is still the dominant payment method for e-commerce in Morocco, but it comes with a costly step: someone has to call every customer to confirm the order before it ships, or risk a wasted delivery.
Sellers were paying for a call-center team to make hundreds of these confirmation calls a day, most of which follow the same short script.
Approach
I built a voice agent that places the call itself: speech-to-text (Whisper), a lightweight understanding layer, and speech synthesis (ElevenLabs) wired together over Twilio telephony.
The hard constraint was latency — a voice agent that pauses awkwardly between turns doesn't feel like a phone call. I tuned the pipeline (streaming STT, short-context prompts, Redis for session state) to keep response time close to human conversation speed.
FastAPI runs the orchestration; the agent handles confirmations, reschedules, and cancellations without a human in the loop for the common cases.
Outcome
The agent now handles the confirmation call end to end for a live e-commerce seller, freeing the call-center step for cases that actually need a human.
Stack
Python, FastAPI, Twilio, Whisper, ElevenLabs, OpenAI, Redis