PromethAI - AI assistant

PromethAI - Refine your decisions
PromethAI - Automate your actions
PromethAI - Get quick recipes

Inspiration

We were inspired by babyAGI project and the recent papers published by Stanford that give Agents personality and memory. In addition to that, works of Tversky and Kahneman and research experience in the field of Cognitive Psychology helped us with deciding that we want to focus on the issues of cognitive load and decision making process. We were wondering how to enable Agents to move from System 1 (automated, heuristical) thinking to System 2 (complex problem solving).

What it does

PromethAI is a Python-based AGI project helps user break down problem into components, navigate to a goal and then supports with executing that goal . We focused on helping users choose meals in the first iteration.

The AI agent now has the ability to suggest meal options based on a user's specified goal, such as a fast meal, a tasty meal, or a healthy meal. It also helps user navigate the decision tree, and in the end automate the actions.

The agent can also suggest meal options based on a user's preferences, such as a user's favorite cuisine or a user's favorite restaurant. The backend also has the ability to retrieve a list of restaurants from Google Maps and suggest matching food options based on the user's preferences. We also provide Zapier integration, so you could send a message to your mom to ask what ingredient for your recipe you might be missing, or block time in calendar for your meal.

Overall, PromethAI is a practical application of AGI technology that has the potential to help users make informed food choices based on their goals and preferences.

How we built it

We built a FastAPI endpoint around Langchain chains that serve JSONs generated by the OpenAI LLM. In this way, we could keep the communication between the frontend and backend flexible.

We use Pinecone to store user summaries and preferences in order to be able to generate better outputs for users down the line. We built Dart/Flutter app that dynamically generates decision points based on user inputs and LLM queries and provides generated results of recipes and other actions to the user.

Challenges we ran into

Latency and caching were some of the biggest issues, but by combining prompt engineering with partial results + caching strategies, we managed to reach a solid response time for the system.

Accomplishments that we're proud of

We are very happy with implementing decision tree logic, and we believe it can be a step forward in future interactions with the AI

What we learned

LLMs and Langchain are still very new, and having a system that scales well is not a trivial task

What's next for PromethAI -

We've added Zapier Actions, and we hope to add functions and more dynamic input options for the user. We also need to do a few user interview and optimise our designs based on that. We plan to continue working on this project in the future.