Caylent Catalysts™
Generative AI Knowledge Base
Learn how to improve customer experience and with custom chatbots powered by generative AI.
Fintech Innovator Accelerates Payment Support Automation with Generative AI on AWS
Generative AI
→ Delivered a browser-based generative AI assistant for loan payment and rescheduling workflows, enabling a natural language voice and chat experience tailored to loan servicing.
→ Integrated AWS services including Transcribe, Bedrock (with Anthropic Claude), Polly, and CloudWatch to support real-time voice processing, secure orchestration, and compliance monitoring.
→ Built a real-time voice interface using a Next.js frontend, WebSockets, and a FastAPI backend—reducing call center load and minimizing IVR escalations through workflow automation.
→ Implemented a deterministic scripting model to tightly control AI responses, maintain regulatory compliance, and build trust in AI-assisted payment interactions.
→ Designed a modular, portable architecture to support future IVR integration and simplify ongoing platform evolution.
Innovating with Generative AI While Prioritizing Trust
In a sector where innovation is often tempered by regulatory complexity, Best Egg stands out for its bold yet thoughtful approach to adopting new technologies. As a leading fintech platform, Best Egg has continually embraced digital transformation to improve how consumers access and manage credit—proving that it's possible to deliver innovation at scale without compromising trust, compliance, or control.
Driven by a mission to "help people feel more confident about their finances," Best Egg combines a fully online model with a strong customer-centric philosophy. The result is a streamlined experience that enables fast personal loan approvals, flexible financial tools, and simple, intuitive interactions. With over a million loans originated since 2014, Best Egg continues to treat technology not just as an enabler, but as a lasting competitive advantage.
Transforming the Self-Service Payment Experience
In the financial services industry, precision, compliance, and trust are non-negotiable. With customer touchpoints being vital to both retention and satisfaction, Best Egg approached Caylent with a vision to explore how generative AI could improve their customer support workflows.
Their collaboration with Caylent was rooted in the goal of improving self-service payment capabilities. Rather than relying solely on rigid Interactive Voice Response (IVR) flows that often lead to customer frustration, they envisioned a natural language experience—one that could process payment requests, reschedule auto-payments, and hand off to live agents when necessary. Caylent delivered a production-ready proof of concept that showed what was possible using AI, giving Best Egg a platform to evaluate and eventually operationalize conversational automation.
This MVP served as a controlled, high-impact demonstration of what generative AI can enable in financial services when implemented responsibly. With the AI assistant, Best Egg can now offer browser-based self-service experiences that understand customer intent in natural language, execute secure payment actions through internal APIs, and ensure a seamless escalation to human support when needed.
The solution reduces friction in common workflows, cuts down on IVR falloff, and allows internal teams to iterate quickly without large-scale infrastructure changes. The same core system also supports a chatbot variant, giving Best Egg flexibility across voice and text channels. And with extensible architecture and fully scripted responses, the platform maintains complete control over compliance while setting the stage for future deployment into their existing IVR systems.
Problem
Best Egg’s existing Interactive Voice Response (IVR) system, powered by Alvaria, supported basic automation but relied on rigid phone tree structures that tend to offer limited and often clunky user experiences that can lead to higher call abandonment rates. They wanted to create a more natural, conversational self-service experience—one that would allow customers to simply say what they needed (“I’d like to make a payment” or “Can I reschedule my autopay?”) and have the system understand and act on it.
While the original goal was full IVR integration, the deeply embedded Alvaria system and limited internal engineering bandwidth made it more viable to pursue a browser based proof-of-concept (PoC) to showcase the capabilities of the system. There were important technical needs the solution would have to meet:
- Secure access to internal APIs for executing payment-related actions
- Deterministic control over assistant responses to meet compliance standards
- Flexible architecture to reduce long-term maintenance burden
- Seamless compatibility with existing backend workflows
- Accurate natural language understanding and intent recognition
- Smart escalation to human agents when needed
This approach offered a faster path to value, eliminated deep coupling with the Alvaria IVR system, and still allowed the AI assistant to interact with real backend APIs and workflows—making the experience both realistic and extensible. It also enabled Best Egg to test conversational AI in a lower-risk environment and gather feedback from stakeholders before committing to a more complex production rollout.
Solution
Caylent delivered a production-ready proof of concept that showcased the utility of generative AI for high-sensitivity customer service workflows—without compromising security, compliance, or user experience. The solution was designed to be flexible, lightweight, and fully compatible with Best Egg’s existing backend infrastructure, while allowing room for future IVR integration.
Natural Language Understanding for Payment Workflows
Technologies used: Amazon Transcribe, Anthropic Claude (via Bedrock)
The assistant was built to interpret voice-based customer requests using natural language, enabling users to bypass rigid phone trees and simply ask for what they needed. Amazon Transcribe converted speech to text in real time, which was then passed to Anthropic Claude via Amazon Bedrock for intent classification and entity extraction. This allowed the assistant to understand user goals—such as making a payment or rescheduling auto-payments—and extract relevant details like amounts or dates.
Importantly, the LLM was not used to generate open-ended responses. Its role was strictly limited to recognizing intent and parsing structured inputs to maintain control and predictability in every interaction.
Secure Execution of Financial Transactions
Technologies used: Internal Best Egg APIs, FastAPI backend
Once user intent was determined, the assistant triggered real-time actions through Best Egg’s internal APIs—originally built to support their IVR platform. This included checking payment eligibility, initiating payment transactions, and updating auto-payment schedules. By leaning on these preexisting APIs, Caylent avoided duplicating business logic and ensured the assistant stayed aligned with Best Egg’s compliance and operational policies.
A Python-based FastAPI backend served as the orchestration layer, managing WebSocket communication between the frontend and AWS services, maintaining state, and securely handling API calls.
Voice and Chat Flexibility with Scripted Responses
Technologies used: Amazon Polly, Pre-defined response scripting library, WebSockets, Next.js
To maintain tight control over bot responses, Caylent implemented a deterministic script library. Each user intent mapped to a curated set of approved responses—no open-ended text generation was allowed. These responses were then converted to speech using Amazon Polly and streamed back to users via WebSocket connections.
The frontend experience was delivered through a modular web application built in Next.js, which supported both voice and chat interactions. For the chat experience, Caylent simply disabled voice-specific components (Polly and Transcribe), allowing Best Egg to use the same logic and workflows across both channels.
Observability and Extensibility
Technologies used: Amazon CloudWatch, modular architecture, prompt configuration
Caylent built comprehensive observability into the system using Amazon CloudWatch. Each session generated detailed logs, capturing transcription results, intent detection, API activity, and overall session flow. These logs supported downstream analytics and troubleshooting.
The architecture was designed for future scalability. Adding a new workflow or intent required minimal changes—primarily updates to the LLM prompt config and scripting library. Caylent also prepared a knowledge transfer session to walk Best Egg through how to extend or adapt the system independently post-handoff.
