Let's Talk
Logo Acerto
Financial
About the Client

Acerto is a Financial company based in Brazil focused on providing debt negotiation services for general customers

Solution Area

Artificial Intelligence

Industry

Financial

Client Location

Brazil

The Challenge

Acerto has implemented a customer service process through the Zendesk platform. This process is designed to handle the high volume of customer support tickets that Acerto receives daily. On average, there are over 1,000 tickets that need to be answered by Acerto’s customer service team. These tickets cover a wide range of topics, including negotiation proposals, requests for second copies of invoices, payment notices, and other customer inquiries. Acerto relied on its team to respond to these tickets. This resulted in a significant amount of manual work for the customer service team. Many of the tasks associated with responding to tickets were repetitive and time-consuming, taking away from the team’s ability to focus on solving more complex problems.

The Business Solution

In order to overcome this business challenge, we proposed a Generative AI agent capable of classifying the tickets according to its proper categorization, generate answers that would meet or solve users requests and escalate to a human assistant in case of a request that could not be solved by the agent.

The Solution

To implement this business solution, Avenue Code’s development team collaborated closely with the client to establish an infrastructure within Google Cloud capable of processing more than 1,000 tickets daily via an API. To achieve this, the team employed Vertex AI and the Gemini-pro model as the core components of the inference pipeline. Additionally, a serverless environment was implemented, leveraging API Gateway and Load Balancer to ensure robust security measures.

1

Vertex AI and Gemini-pro Model:

- Vertex AI serves as the backbone of the inference pipeline, providing a comprehensive suite of machine learning tools and services.

- The Gemini-pro model, specifically designed for natural language processing tasks, plays a crucial role in classifying and summarizing ticket subjects based on their descriptions.

2

Serverless Environment:

- Serverless architecture was adopted to eliminate the need for managing and maintaining servers, enhancing scalability and cost-effectiveness.

- API Gateway facilitates secure and scalable handling of API requests, ensuring efficient communication between different components.

- Load Balancer distributes incoming traffic across multiple instances, promoting high availability and preventing single points of failure.

3

Integrated API for Zendesk Connection:

- A secure API was developed to establish a seamless connection with Zendesk.

- This API enables the exchange of data between the custom solution and Zendesk, ensuring real-time synchronization of ticket information.

4

Serverless Solution for Request Handling:

- A serverless solution was implemented to handle the high volume of requests efficiently.

- Serverless functions scale automatically based on demand, optimizing resource utilization and minimizing costs.

5

RAG Solution with Cloud SQL Database:

- A RAG solution was deployed using Cloud SQL's database, to make sure the agent is able to answer factual questions and produce content that requires specific knowledge.

Leveraging the power of the AI Studio Platform, a comprehensive suite of tools designed to streamline the creation of software services enhanced by Generative AI, we were able to dramatically accelerate the development process. The platform’s integrated agent development environment and advanced capabilities minimized the need for boilerplate code, allowing us to focus on delivering high-value features. By utilizing the AI Studio Platform, we completed the entire solution within 14 working days, a task that traditionally would have required at least 30 extra days of development.

The Results

LLM Performance:

Leveraged GCP's advanced LLM features, such as AutoML and Cloud TPU, to significantly enhance the performance of our machine learning models. This resulted in improved accuracy, faster training times, and reduced computational costs.

Cost Reduction:

Achieved a 70% reduction in operational costs by deploying our application on GCP's serverless platform. This eliminated the need for costly on-premises infrastructure and allowed us to scale our application dynamically based on demand.

Data Processing:

Optimized more than 90% of the workload by utilizing GCP's powerful data processing tools, such as BigQuery and Dataflow. This enabled us to handle large amounts of data efficiently and cost-effectively.

Security Enhancements:

Implemented robust security measures to protect our application and data, including automatic VPC, DNS, and IP protection. Additionally, we utilized API Gateway and Load Balancer to control access to our application and distribute traffic securely.

Key Metrics 

Over 1.000 tickets processed daily

Accuracy of 90% over 25 topics

Reduction of 70% in operational costs

Optimization of over 90% of the workload

Tech Stacks 

CloudSQL

GCP Artifact Registry

Serverless

Google CloudRun