Multi-Model AI Pipeline for Customer Analytics in a Contact Center SaaS

About Our Client

The Client is a North American provider of all-in-one contact center software (voice, email, SMS, chat, social media, and other channels) with more than 30 years of market presence. Its SaaS product is widely adopted by large enterprises, upper mid-market organizations, and public services, helping deliver consistent, frictionless customer experiences.

Making AI Work Inside Enterprise-Grade Contact Center Environments

Inspired by the rapid advancement of AI technologies, the Client wanted to explore how intelligent capabilities could enhance its contact center platform and make the product more valuable for customer service teams. At the same time, the platform had to remain stable and scalable: any AI enhancement needed to integrate seamlessly into the existing agent desktop application without disrupting established workflows or requiring the Client’s customers to rebuild their environments. In addition, the new capabilities needed to support large-scale enterprise deployments and remain compatible with cloud, on-premises, and hybrid infrastructures already used by the Client’s customers.

Having partnered with ScienceSoft for several years on platform modernization initiatives, the Client relied on the team’s deep knowledge of the product architecture and expertise in applied AI.

Engineering Real-Time AI for Enterprise Contact Center Platforms

Using the existing knowledge from previous collaborations, ScienceSoft’s team immediately identified a technically and commercially sustainable opportunity to add an AI-powered speech and text analytics layer to the SaaS platform. ScienceSoft’s proposed to build the AI layer as an independent module integrated via an API gateway, which would allow the Client to extend the existing platform without tightly coupling new AI functions to the core SaaS architecture. It also relied on AI models to avoid unnecessary complexity, vendor lock-in, and dependence on costly external AI services. The approach resonated strongly with the Client, who decided to move forward with the initiative together with ScienceSoft.

Following a joint discovery phase, ScienceSoft and the Client collaboratively defined a focused set of AI capabilities that were easy to implement and would bring immediate value to the contact center platform. The project scope covered four core AI functions: real-time call transcription, automated conversation summarization, sentiment analysis, and contact reason detection.

Real-time transcription layer

ScienceSoft’s team applied a dual-model strategy to the speech-to-text layer to give the Client flexibility across different deployment scenarios, infrastructure constraints, and other operational requirements.

For cloud-based deployments, ScienceSoft integrated Speechmatics as a managed automatic speech recognition (ASR) service. The choice was driven by its consistently strong transcription accuracy across multilingual conversational audio, operational maturity in production environments, and ability to handle real-world contact center audio conditions (accents, noise, and speaker interruptions) without additional tuning.

For self-hosted and cost-sensitive environments, ScienceSoft implemented NVIDIA Parakeet TDT 0.6B v3. Built on a FastConformer-based architecture, the model provides low-latency, high-throughput transcription and can be deployed on moderate GPU infrastructure without relying on vendor-controlled cloud-based ASR services.

Customer insights layer

After enabling real-time call transcription, ScienceSoft’s team built a natural language processing (NLP) layer that transforms raw transcripts into structured operational and customer insights.

For conversation summarization, ScienceSoft integrated Mistral AI’s Mistral NeMo Instruct 2407. ScienceSoft selected the model for its reliable instruction-following capabilities and support for large context windows, which allowed the system to summarize long customer conversations while reducing the need for aggressive transcript chunking that could disrupt conversational continuity. To keep infrastructure costs under control, ScienceSoft deployed the model through llama.cpp in GGUF format, enabling efficient inference across both CPU- and GPU-based environments.

For contact reason detection, ScienceSoft implemented GLiClass base v3.0 to support dynamic zero-shot classification. Instead of relying on a fixed set of pre-trained intent categories, the system allows support teams to update contact reason taxonomies without retraining the model whenever products, workflows, or support scenarios change. ScienceSoft also enabled multi-label classification, allowing the platform to identify multiple contact reasons within a single customer interaction.

For sentiment analysis, ScienceSoft integrated a multilingual sentiment model from Tabularis AI to continuously evaluate customer emotions, or sentiment, throughout conversations. He prioritized low inference latency and lightweight execution so the sentiment layer could process streaming text in near real time without creating significant processing bottlenecks.

ScienceSoft structured the solution as a set of independently deployable microservices, each responsible for a specific stage of the AI pipeline. This allowed the Client to isolate compute-intensive workloads, scale each function independently, and simplify long-term model upgrades and replacements.

ScienceSoft chose AWS as the main deployment platform since it was already widely adopted within the Client’s product environment. Each service was containerized using Docker and deployed on Amazon ECS, with AWS Fargate used to reduce infrastructure management overhead and enable automatic scaling based on demand. An AWS Application Load Balancer provided secure API exposure and path-based routing across AI services.

Value-Adding and Cost-Effective AI Modules for a Contact Center SaaS

ScienceSoft delivered a production-ready AI layer for the Client’s contact center SaaS platform, consolidating call transcription, conversation summarization, sentiment analysis, and contact reason detection into a unified real-time processing pipeline.

As conversations unfolded, the system continuously transformed live audio streams into structured insights inside the agent desktop app, giving support teams immediate visibility into customer intent, sentiment shifts, and conversation context during active calls. Agents no longer needed to manually capture key discussion points or reconstruct conversations for CRM updates, which reduced after-call workload and improved the consistency of customer interaction data. At the same time, supervisors and analytics teams gained faster access to operational insights for service quality monitoring, trend analysis, and workflow optimization.

From a technical perspective, the Client received a modular, cloud-native AI architecture that allows individual AI capabilities to scale independently based on workload demands. The solution was designed to work consistently across cloud, on-premises, and hybrid deployments while enabling future AI models and capabilities to be introduced without major changes to the platform’s core architecture.

By building the solution on top of the Client’s existing AWS-based infrastructure and leveraging cost-efficient AI technologies, ScienceSoft helped the Client introduce advanced AI functionality without high operational costs, restrictive model licensing, or dependency on closed vendor-controlled AI ecosystems.

Technologies and Tools

Python, Rust, NVIDIA Parakeet TDT 0.6B v3, Speechmatics, Mistral-Nemo-Instruct-2407 (GGUF via llama.cpp), Tabularis AI multilingual sentiment model, GLiClass-base-v3.0, Docker, Amazon ECS, AWS Fargate, AWS Application Load Balancer.