FoodNav

This is our initial MVP for AI and Food Insecurity Case Competition by Robert H. Smith School of Business, University of Maryland. We have developed an AI based solution which uses a combination of advanced NLU, LLMs(Generative Ressponses) and a multilingual agent to help people find food assistance resources in the DMV area.

Our data is from Capital Area Food Bank, which is a non-profit organization that provides food assistance to people in need in the DMV area.

Installation:

Refer to the README for installation instructions.

Features:

Find’s the top 3 nearest food assistance resources based on the user’s location and takes into account the user’s preferences and needs.
We use geocoding api from google maps to geoencode users address and geoencoded addresses of the resources to calculate the distance.
We use Dialogflow CX to create a multilingual agent that can understand and respond to user queries in multiple languages. (Currently only English and Spanish are supported)
Primarily we focused on the Call based aspect of the agent, but we also have a web app and SMS based solution.
We decided that generative responses and decision making are not the best use case for LLMs, so we used Intents and Entities to create a decision tree based agent using Dialogflow CX.
Also, supports call companion feature, so the user can chat with the agent while on the call.
After the user ends the call, we send the user a multilingual feedback form to get feedback on the agent’s performance.
Have support for human agent handoff in case the user is not satisfied with the agent’s response.
On the server since the database was less than 1MB, we used SQLite and pandas to store the data and perform queries. We also used FastAPI to create a REST API. This was the webhook for the Dialogflow CX agent.
Also have APIs for the maps part of the web app.
We used test cases and the analytics dashboard to monitor the performance of the agent and make improvements.
For the web app we also made sure to use i18n to support multiple languages.
We used ngrok to expose the local server to the internet for testing purposes.

View the Project

Collaborators:

@worldsoldure - Team Member, Primarily worked on the maps part of the web app, user research and presentation.
@eyepatch0 - Team Lead, worked on the Dialogflow CX agent, server and the web app.

Contact:

Feel free to reach out to me on LinkedIn if you have any questions or feedback. I’m always open to suggestions and ideas for improvement.