As a category leader in the call center automation space, Vodex has created a reputation as one of the best AI-powered outbound call making services in India. This is underscored by the more than $2 Million they have raised from marquee investors such as Unicorn India Ventures and Pentathalon Ventures.

With call center operations, the latency between the user and the bot has to be as low as possible to ensure retention and provide a human-like experience. As Vodex grew, this latency started to increase, along with the compute costs associated with delivering the experience.

A model ensemble such as a voicebot necessitates an understanding of model orchestration along with optimized models that are highly performant to realize a human-like interaction. Responses have to be prompt and include nuances like tonality and inflection to build user trust and minimize churn.

Vodex faced some key challenges:

• A high user latency of 2.4s was leading to customer drop-offs and dissatisfaction. This great of a delay between query and response led to the calls feeling more robotic
• A sub-optimal model ensemble was causing not just this increased latency, but also leading to increased costs as the compute resources were not being utilized efficiently.

Simplismart completed a comprehensive analysis of Vodex’s load profile and use case, before implementing a solution involving an ensemble of 3 open-source models, Whisper by OpenAI, Llama by Meta, and Tortoise TTS by James Betker, all deployed on the Simplismart platform.

• Simplismart optimizes models and deployments at multiple levels including the model-chip interface and the serving layer allowing for not just faster model performance but also superior ensemble performance leading to extensive cost savings of over $100,00 till date.
• Simplismart optimized models allowed for a transformation in the workflow, improving transcription accuracy and bot responses that are more intelligent and human-like.