Syncode client delivery
Conversational AI in production
INPUTDevice session
APPROACHSpeech + inference
OUTCOMESpoken response
OBSERVE · LEARN · IMPROVE
The problem
Deliver a device voice experience across speech recognition, language-model serving and speech synthesis, while preparing for seasonal demand.
The approach
Self-hosted vLLM, Whisper and Piper serving; typed device sessions, Kubernetes services, and connected traces, logs and metrics.
The result
Delivered and operated through a seasonal demand peak. Built traffic forecasting and capacity-calibration tools that turn observed demand into load-test scenarios; resolved inference-memory and telemetry incidents.
A closer look at the engineering
Capacity engineering started with observed traffic: analyse demand, fit curves, calibrate serving capacity and generate load scenarios. The output is a load-test scenario for reviewing serving capacity before peak demand.