FreeFoodForecastFilterForFrugalFoodFestivityFollowingFriends

Extraction Results
Query Options
App Modes
Morning Brief

Inspiration

We are college students who have always been passionate about technology and helping others. As we brainstormed ideas, we thought about the struggles that many college students face, including ourselves. We often find ourselves skipping meals or eating cheap, unhealthy food because we can't afford anything better.

We decided to create an app that would help students find free food on and around campus. We spent long hours coding and designing the app, fueled by our passion to help others and the hope that our app could make a difference.

What it does

Core Features/Modes

Automatically reading and parsing Emails for event invitations, creating corresponding Calendar Events
Functionality for only processing events that offer free food
Morning Brief: create inbox summary

Interface

Flexible Query Options: gives users control over the queried emails' senders/subjects/labels and more
Extraction Interpretability: a major drawback to leveraging LLM's general purpose capabilities is the lack of interpretability. It's unclear why LLMs infer the way they do. We added a mitigating feature of "source" strings where we highlight the text spans the LLM supposedly attended to when extracting a feature. Although this is a post-hoc explanation, we assert it still provides information about the final inference.

We hope these functionalities help students find sustenance, make more of the varied event offerings at University, and also reclaim precious time that's otherwise wasted pouring through emails.

How we built it

We leverage Google's SDK APIs for Gmail and Google Calendar to read emails and write events to Calendar. We leverage large language model's in-context learning capabilities for both classification and information extraction tasks. The front-end is created with the streamlit framework.

Challenges we ran into

Large language model's opacity makes analytically optimizing in-context inputs impossible, forcing us to depend on inefficient black-box optimization techniques. Furthermore, certain templates are effective for different inputs with unclear distinction.

Accomplishments that we're proud of

successfully connecting a complete pipeline with a number of components
creating an elegant front-end
an amazing logo and project name

What's next for FreeFoodForecastFilterForFrugalFoodFestivityFollowingFriends

Further Free Food Finding

Also further prompt optimizations. We're investigating smaller, but similarly promising LLMs to deploy ourselves to drive down latency. We're also looking to enrich the Morning Brief mode with structured information in addition to the brief.