Inspiration

Our idea stemmed from looking at the sponsor challenges: when we saw 11ElevenLabs' voice tools, we tried to picture possible projects that integrate text-to-speech and speech-to-text with LLMs in an interesting and entertaining way.

What it does

The project puts players in the shoes of a insurance claimant with questionable morals. The primary goal is to get the insurance representatives to accept your claim, no matter how much you may need to stretch the truth. Players can choose from various scenarios and characters, offering a wide range of experiences. Each character has a distinct personality, set of quirks, and backstory.

How we built it

Given the span of the project, Gemini agents were used to create a large portion of the codebase. This allowed us to use a variety of libraries and, consequently, implement a diverse set of features that would not have been possible otherwise. In addition, templates were created to ensure that the prompts fed to the LLM were reliable, protected against malicious users, and conducive to organic responses. This approach ultimately enabled a shift from focusing on pure logistics to prioritizing big ideas.

Challenges we ran into

  • Latency between user speech and LLM response
  • Creating fair and consistent characters
  • Avoiding the development of game metas
  • Building scenarios and gauging their difficulty

Accomplishments that we're proud of

Above all, we are most pleased with our central idea. Namely, making use of machine learning allows for game ideas that aren't possible with traditional rules based programming. Moreover we are proud of:

  • The connection between user speech and the LLM
  • Voice filters that customize the speech of each character

What we learned

  • Creating an AI pipeline that integrates multiple types of models
  • Prompt engineering
  • Selecting the right libraries for the job

What's next for Insurance

  • Implementing more scenarios and characters to further flesh out the game
  • Refining UI and UX to improve navigation
  • Creating a system that allows user progression and saving data
  • Improving character memory and coherence throughout a conversation

Built With

Share this project:

Updates