Tuali Agent

What inspired us?

We were inspired by the daily reality of small business owners and local vendors who repeat the same purchasing routines over and over, often while juggling customers, inventory, family responsibilities, and limited time. We wanted to build something that feels less like another app to manage and more like a helpful assistant: one that remembers what a business usually buys, suggests what they may need next, and makes restocking faster.

What does it do?

Tuali Agent is a smart ordering assistant that helps users build repeat purchases quickly. It analyzes a customer’s purchase history and current cart to generate a “Pedido Inteligente,” or suggested order, along with personalized product recommendations.

Users can review suggested items, adjust quantities, add recommendations to their cart, confirm an order, and rate how useful the recommendations were. That feedback is saved so the system can keep learning and improve future suggestions.

The big challenge: building it

The biggest challenge was connecting a polished mobile shopping experience with an intelligent backend. We built a Flutter app for login, dashboard browsing, smart orders, cart management, and order confirmation, then connected it to a FastAPI backend that uses customer order history, product embeddings, pgvector similarity search, and reinforcement learning feedback.

The real work was making the AI feel useful inside a normal shopping flow, not like a separate feature bolted on at the end.

What did we run into?

We ran into challenges around recommendation quality, duplicate products, and keeping the app responsive when the backend was unavailable. Some catalog items were semantically very similar, so we had to deduplicate recommendations carefully instead of showing the user repeated versions of the same product. We also added local fallback data so the app could still demonstrate the experience if the API failed.

Another challenge was turning user feedback into a useful learning signal. We implemented a lightweight rating flow that stores reward data for future reinforcement learning training.

Accomplishments that we're proud of

We are proud that Tuali Agent works as an end-to-end product, not just a concept. It has a mobile interface, persistent cart, dynamic smart orders, personalized suggestions, backend order saving, and a feedback loop for improving recommendations.

We are also proud of making the experience feel simple: the user does not need to understand AI, embeddings, or reinforcement learning. They just see a faster, more personalized way to restock.

What we learned throughout this challenge

We learned how important it is to design AI around real user behavior. A recommendation system is not only about returning similar products; it also has to avoid repetition, respect what the customer already buys, adapt to the current cart, and fit naturally into the purchasing flow.

We also learned how valuable feedback loops are. Even a simple star rating can become the foundation for a system that learns which recommendation strategies work best for different customers and contexts.

What's next for Tuali Agent?

Next, we want to make Tuali Agent more conversational and more adaptive. We would like to connect the chatbot experience directly to the ordering flow, support voice-based ordering, improve the reinforcement learning model with more real usage data, and add smarter strategies like stock replenishment reminders, promotions, personalized combos, and volume upsells.

Long term, Tuali Agent could become a practical AI companion for small businesses: helping them save time, reduce missed inventory, and make better purchasing decisions with less effort.

Built With

• fastapi
• flutter
• langchain
• openai
• pgvector
• postgresql
• rag