The problem Quivo solves

Online sellers encounter challenges in effectively presenting their products on ecommerce platforms, making it difficult to convey their true value to potential customers.
The absence of human-like interactions in the online shopping experience hinders trust building and impacts sales.
Our goal is to bridge the "human gap" between buyers and sellers in the digital realm using AI agents to enhance trust and communication, ultimately increasing online sales and customer satisfaction.
Enhanced User Experience: Quivo can significantly improve the user experience for both online sellers and buyers. Sellers can benefit from AI-driven negotiation support and expert advice, while buyers can enjoy more personalized and informative interactions.
It has a usability impact of,
Increased Trust: By introducing human-like interactions in the online shopping experience, Quivo can help build trust between buyers and sellers. Trust is crucial in e-commerce, and it can lead to higher sales and customer loyalty.
Boosted Sales: Online sellers can leverage Quivo's capabilities to strategically promote and sell their products. This can result in increased sales and revenue for e-commerce businesses, contributing to economic growth.
Empowering Small Businesses: Quivo's B2B focus can be particularly beneficial for small online businesses that may lack the resources for extensive customer support. It levels the playing field by providing them with advanced AI tools.
Accessibility and Inclusivity: Quivo can make online shopping more accessible to a wider audience, including people with disabilities or those who may require personalized assistance during the shopping process.
Social Inclusivity: Making e-commerce more accessible and user-friendly can benefit individuals who might face barriers to in-person shopping, including those with mobility issues, busy schedules, or social anxiety.

Challenges we ran into

Fine-tuning a Language Model (LLM) for Indian negotiation presented an array of challenges that demanded creative problem-solving and resourcefulness. These challenges encompassed computing resources, data collection, optimization techniques, model size, and integration between the frontend and backend.

First and foremost, computing resources posed a considerable challenge. Our fine-tuning efforts were constrained by the limitations of a single 16GB RAM GPU, which restricted the model's training capacity. The modest GPU capacity necessitated careful optimization techniques to ensure efficient model training and performance. We utilized methods like LoRA (Low-Rank Adaptation) to maximize the utilization of available resources and fine-tune our LLM effectively.

Data collection for negotiation was another complex task. Gathering sufficient, high-quality data specific to Indian negotiation was a significant hurdle. This required rigorous curation and annotation of negotiation dialogues to train the model effectively. Ensuring data diversity and relevance was crucial to equip the LLM with the necessary knowledge for accurate responses.

Moreover, integrating the LLM without deploying it on cloud computing infrastructure was a priority to optimize cost efficiency. Deploying a model on the cloud can be costly, and to save on expenses, we explored alternatives. We devised solutions that allowed us to use the model in an on-premises or local server environment without relying on expensive cloud resources.

Model size was yet another challenge to tackle. Our base LLM model was originally designed for a broader range of tasks, and its size made it impractical for real-time negotiation scenarios. We had to fine-tune and adapt the model to be more streamlined and focused on the specific needs of Indian negotiation, ensuring it could run efficiently on our available hardware.