ReMi | Devpost

example of usage

ReMi is a revolutionary solution designed to streamline your grocery shopping experience by transforming unstructured ingredient lists into optimized, organized grocery lists. Our fine-tuned industrial model intelligently groups items by store sections, significantly speeding up your shopping trips.

Inspiration

Our project is inspired by the common challenge of spending excessive time in grocery stores. We love cooking and eating healthy homemade food, but we don't enjoy the inefficiency of navigating through the store multiple times to find every item on our list. Typically, when you compile ingredient lists from multiple recipes, you end up with a chaotic, unorganized list that forces you to zigzag through the store.

Solution

ReMi organizes these unstructured shopping lists by store sections and items, ensuring you only need to visit each section once. This not only saves time but also makes the shopping experience more pleasant and efficient.

Challenges and Innovations

One of the most interesting challenges we faced was tuning the embedding models. We needed specialized embeddings that go beyond casual semantic similarity. For example, while semantic models might group "cucumber" and "pickled cucumber" together, these items are located in different store sections. Our model needed to reflect such practical distinctions, grouping fresh vegetables together. By leveraging contrastive learning, we fine-tuned our model, making it two times faster and using half the memory. This led to a 12% improvement in classification accuracy for our clusters—a significant accomplishment in a novel task of embedding utilization.

Accomplishments

Our major achievement lies in the fine-tuning of the Mistral model. This innovative approach has not only enhanced the accuracy but also the efficiency of the shopping list optimization process. Additionally, our app is user-friendly and accessible on both PCs and smartphones, allowing users to utilize it directly in supermarkets. It's versatile, supporting all types of devices.

What we've learned

Through this project, we have deepened our expertise in working with large language models and tuning them effectively. We utilized a range of resources, including Nebius GPUs, Weights & Biases, and Hugging Face platforms, which were instrumental in our development process.

Future Plans

Moving forward, we aim to enhance ReMi further by incorporating store catalogs, allowing users to select their preferred store for optimized shopping lists tailored to that store’s layout. We also envision automating the process of extracting ingredient lists from recipes, so users won't even need to create a shopping list manually.

We are thrilled about the potential of ReMi and look forward to continuing our journey in making grocery shopping a seamless and enjoyable experience.