Verizon x J Balvin

We recognized the excitement and responsibility of bringing a rich, personalized interaction with a large language model to millions of viewers. This would be many users' first experience interacting with AI, and we set out to create a fun, dynamic experience that was safe and responsible when addressing any topic that the general public would throw at it.

We also had to wed an agile methodology with a more traditional production timeline and navigate the challenge inherent in non-deterministic AI models (meaning they are likely to have different results every time) to embody a unique personality and stay on brand for both J Balvin and Verizon. And...we only had 5 weeks to go from concept to prototype through regulatory requirements and into development, testing, and going live during the Super Bowl.

We needed to develop an interactive experience that would be intelligent, fluent in both Spanish and English, and embody the vibrant tone, style, and energy of José, a.k.a. J Balvin. Working with the J Balvin team, Colombian and American copywriters, and delving into his music, life, and communication style, we were able to engineer prompts and filters to inform the AI bot and allow a wide-ranging, yet on-topic conversation.

By iteratively conditioning and fine-tuning the model, we were able to develop an authentic chat experience that could speak about any subject as well as intelligently deflect away from sensitive and inappropriate topics.

Our team developed a web application prototype to allow the Quality Assurance team and stakeholders to efficiently test the application behavior and evaluate multiple versions of the fine-tuned GPT model. This prototype replicated the SMS user experience with the added advantage of a graphical user interface, significantly enhancing our capability to test and manage multiple models and versions.

One of the significant challenges in the project was managing the possibility of users engaging in both Spanish and English. The application could detect whether the conversation was happening in English, Spanish, or a blend of both, and then create the right prompt for GPT-4 to answer in the language the user expected.

Understanding the non-deterministic nature of GPT-4, we added an extra layer of filtering and set up predefined responses. This move was essential in tackling problems like profanity and off-limits topics, making sure the model's outputs stayed clean and met user expectations.

The Super Bowl campaign captured the attention of hundreds of thousands of users, setting the stage for spikes in traffic as the campaign went live. Unlike typical applications that see a steady rise in traffic, our app had to be ready to handle immediate spikes in demand. To prepare, we designed a series of stress tests to simulate peak load conditions, allowing us to identify and address any performance bottlenecks and unexpected behavior. Additionally, we setup an infrastructure capable of managing a significant number of concurrent requests, designed to scale up swiftly in response to a sudden influx of users.

On Super Bowl day, Antfood, The Community and Verizon coalesced together to form the "war room". A tactical team was put into place to track real-time user engagement, monitor any foul play and safeguard against any technical issues.

At the start of the 3rd quarter, Verizon aired a :60 commercial featuring J Balvin on Univision capturing the attention of millions. J Balvin's phone number appeared several times during the commercial prompting thousands of fans to text the number. J Balvin's phone was lit...

The result was a charming and fun experience utilizing novel technologies to engage and delight fans and prove that Verizon's network doesn't break.