Prompt2Cart

UPDATE ALL OF THIS LATER AFTER THE PROJECT IS DONE

Inspiration

We wanted to eliminate the friction between "thinking" of a product and "buying" it. Traditional e-commerce requires navigating menus, filtering search results, and clicking through multiple pages. We imagined a "Jarvis-like" experience for shopping where you simply state your intent, whether it's "I need a snowboard for an intermediate rider" or "Find me a gift for a coffee lover", and the system handles the rest, all the way to the shopping cart.

What it does

Trovato is an AI-powered shopping assistant that transforms natural language requests into ready-to-buy shopping carts.

Natural Interaction: Users can type complex queries describing exactly what they need.
Intelligent Discovery: The system analyzes the intent and searches across Shopify stores to find the best-matching products.
Agentic Carting: Instead of just showing links, Trovato acts as an agent. It autonomously efficiently "hacks" the checkout flow by securing the necessary access tokens and directly injecting items into a Shopify cart via the Storefront API.
Instant Checkout: The user is presented with a pre-filled checkout link, reducing the path to purchase to a single click.

How we built it

LangGraph & MCP: The backbone of our agent architecture is built on LangGraph. It orchestrates the stateful decision-making loops. We adopted the Model Context Protocol (MCP) to standardize how our agents connect to external tools (like the Shopify Storefront API), making our tools modular and easily extensible.

Unified Commerce Platform (UCP): We built on top of the new UCP to unify our commerce data and interactions, giving us a robust foundation for handling cross-store data.

Groq Inference: We use Groq's Llama 3.1 8B Instant model for ultra-fast agentic reasoning. Its low-latency inference ensures the chat-to-cart loop feels immediate and responsive.

Personalization with MongoDB: We store user profiles (style preferences, sizes, budget, search history) in MongoDB. This data enriches agent queries with personalized context, for example, if a user prefers "streetwear" and "budget-friendly" items, those preferences are injected into the search prompt to surface more relevant products.

Frontend: Built with Next.js 16, React 19, and Tailwind CSS v4 for a glassmorphic, high-performance UI.

Challenges we ran into

Agentic Cart Creation: The biggest technical hurdle was getting the agents to work and accurately put items in the cart. Seamlessly interacting with third-party store APIs to execute cartCreate mutations required precise token handling and validation.

Bleeding Edge Tech: implementing UCP was a challenge as the platform is so new; documentation was scarce, and we had to rely on trial-and-error to integrate it correctly with our agent workflows.

Latency: Ensuring the "chat-to-cart" loop felt instantaneous was critical. We initially experimented with Cerebras for high-speed inference, but ran out of API tokens during development. We pivoted to Groq (Llama 3.1 8B Instant), which provided the fast inference speeds we needed to keep the conversational experience snappy.

Accomplishments that we're proud of

Reverse-Engineering the Checkout Flow: We're particularly proud of the backend logic that can autonomously find a store's public access token and generate a valid checkout URL programmatically.

Seamless User Experience: Achieving the "Speak it. Cart it." flow where a voice/text command results in a tangible cart is a magical moment.

Full Agentic Architecture: Successfully implementing a LangGraph + MCP architecture that actually performs actions (carting) rather than just chatting.

What we learned

The Power of Speed: We learned that for conversational commerce, inference latency is everything. Using Cerebras taught us that faster models lead to better UX.

Standardization Matters: Adopting MCP (Model Context Protocol) showed us how much easier it is to build complex agents when tool interfaces are standardized.

Commerce Complexity: Working with UCP gave us a deep appreciation for the complexity of unifying commerce data across different schemas and stores.

What's next for Trovato

Virtual Try-On with Gemini Image Generation: We want to integrate the Gemini API's image generation capabilities to create a virtual try-on experience. Users will be able to see AI-generated images of themselves wearing clothing items or using products directly within the chat interface. This would transform product discovery from abstract to tangible—imagine asking "show me in that jacket" and seeing a photorealistic preview before adding to cart.

Expanded Personalization: Leverage deeper profile data (body measurements, color preferences) and purchase history to further refine recommendations.

Multi-Store Checkout Optimization: Combine items from multiple Shopify stores into a single optimized checkout flow, minimizing shipping costs and delivery times.

Voice Interface with Elevan Labs: Extend the "Speak it. Cart it." vision with full voice-to-cart capabilities using speech-to-text integration.