Introducing Ocean-1: First Contact Center Foundation Model

We’ve created Ocean-1, a foundation model for the contact center. This large language model is the culmination of our experience in deploying generative AI systems for large enterprises and signifies our latest milestone in advancing the cutting edge AI technology for customer facing conversations.

Powered by foundation models

In our previous post, we shared our thoughts on the Emerging Technology Stack of Generative AI, where we highlighted the distinction between LLMs, AI Systems and AI Products. LLMs (Large Language Models) are the core foundation models that power our suite of products. We believe that a vertically integrated approach to LLM powered products could enable continuous learning and deliver better business outcomes.
Ever since Cresta’s founding in 2017, the same year when Transformer architecture was invented1, we’ve seen a paradigm shift in natural language processing every 2-3 years. So too has our AI stack. In 2018, Cresta was the first company to deploy generative AI in production leveraging GPT-2 Transformer architecture. In 2019, we developed a new approach to direct GPT-2’s text generation to follow behaviors that drive conversation outcomes.2 Ever since, our pace of innovation has only picked up and generative AI today is deployed in all parts of our product suite, from conversation insights to real-time agent assist to virtual agents. For example, our recent blog post shows how to use ChatGPT to diagnose revenue opportunities in contact center.
Cresta's 100x Contact Center
While we continue to work with our partners to bring capable models like OpenAI’s GPT-4 and Anthropic’s Claude to enterprise, we would like to provide world’s leading organizations with best security, privacy and the ability to continuously learn from feedback on private datasets. As a result, we created Ocean-1 by combining powerful base foundation models such as GPT-4 with contact center domain data, using alignment techniques such as self-instruct and reinforcement learning from human feedback (RLHF).
Figure 1


Figure 2
Ocean-1 starts off with a base foundation model and trained based on contact center expertise data. It is designed to achieve three goals:

  1. Better out-of-the-box capabilities for contact center.
  2. Better instruction following by learning through feedback data.
  3. Reduced latency and cost of serving real-time applications at scale.

Out-of-the-box capabilities for contact centers

Enterprise transformation is a journey, not a sprint. While Cresta meets customers where they are and brings them onto a journey to adopt AI, we also pride ourselves on bringing value to an organization on day one. For example, our Value Discovery Assessment relies on out-of-the-box AI models to rapidly diagnose team performances and reasons for outcomes. Here, our foundation model brings fast time to value through a set of out-of-the-box capabilities.



(system message) Decide whether the intent of agent in these messages is Discover Household, where sales agents ask questions to find out how many people and who will be protected by a home security system.

(user message) “K. Now let’s get everything opened up here. Antonio, if you don’t mind me asking, what’s the reason why you’re looking for your home security?

GPT-3 (davinci003) GPT-3.5 GPT-4 Ocean-1
Yes, the intent of the agent in this message is Discover Household. The agent is asking Antonio why he is looking for a home security system, which is a question that would be asked to discover how many people and who will be protected by the system. Yes, the intent of the agent in this message is Discover Household. The agent is asking a question to find out the reason why Antonio is looking for a home security system, which will help the agent understand how many people and who will be protected by the system. Yes, the intent of the agent in this message is Discover Household. No. Discover Household is not the intent of the agent in the given message. The agent is asking for the reason why the customer is looking for a home security system. This is not related to discovering the number of people in the household or who will be protected by the home security system, so the intent is not Discover Household.

Better instruction following through feedback data

In our previous blog post, we discussed the value of feedback data and how a vertically integrated stack could continuously improve model performance. Once our generative AI-powered platform goes live with a customer, our products are designed to capture human feedback on model predictions.
Figure 3
A paradigm seems to emerge about how to improve model’s performance with such data:

  • Zero-shot: describe how to perform the task
  • Few-shot: when zero-shot runs into failure cases, we can add examples to the prompt to improve the model’s accuracy
  • Fine-tuning: When we collect enough failure cases and build a dataset, we can fine-tune the model to achieve even higher accuracy

Although we typically can fine-tune models on feedback data, the effort required to deliver such models is large. To improve the model’s performance on zero-shot and few-shot learning, we rely on Ocean-1 to develop better instruction-following capabilities. As a result, our model can be tuned to hallucinate less on contact center tasks.

Chat Context GPT-4 Ocean-1 Hallucination Note
(customer is looking for a flight change)
Yes, I have the information to make the change. I will update the date of birth to XXX. Please give me a moment to make the update. Sure. May I have your name, email and phone number, please? In this case, the correct agent behavior is to follow the discovery flow, rather than “hallucinating” the outcome.
(customer is looking for refund for a baggage)
Please provide me with the following information:

  1. The original flight details (date, flight number, and route)
  2. The new flight details (date, flight number, and route)
  3. The amount paid for the baggage fees on both flights

Once I have this information, I’ll be able to process your refund request.

I apologize for the inconvenience. May I have the confirmation number to get the process started? Only confirmation number is needed. GPT-4 hallucinates the other information needed.

Faster generation at scale

Our model is optimized for the contact center domain and therefore can be served more efficiently. This is very important for real-time latency-sensitive applications such as Agent Assist, or when multiple LLMs are chained together to deliver the final output (see our previous blog post).

Multiple LLMs

Cresta Opera: No-code AI platform

Cresta customers can interact with these advanced language modeling capabilities through Cresta Opera, our no-code Generative AI platform. As we’ve seen, LLMs are used across our products, from Real-Time Agent Assist to Post-Call Solutions. Opera not only offers a set of out-of-the-box AI models, but also enables users to create models for their specific use cases by leveraging LLM’s instruction following capabilities.
Schedule a demo to see Cresta Opera in action.

Future of contact center foundation models

This is our first step towards the future of contact center foundation models. We believe large language models trained on public internet data provide a powerful basis, but proper instruction finetuning and reinforcement learning unlocks the true potential of such models in contact center domain. This allows us to develop systems and products that can quickly deliver value to enterprise customers and continuously improve through feedback data.

1 – Attention is All You Need
2 – Action Directed GPT-2

Does One Large Language Model Fit All?


The Emerging Stack of Generative AI


How to Use ChatGPT to Diagnose Revenue Opportunities