Restaurar

About the project

Restaurar is a platform in development where people describe situations in their relationships and receive AI-generated guidance. Responses draw on a previously indexed library of books, videos, live sessions and training materials.

Problem

Language models can produce plausible answers without using the sources selected for a product. In a sensitive area such as relationships, the application also needs to handle reports of risk before generating guidance.

Solution

The workflow combines content retrieval with a risk check before generation.

Source material is divided into chunks and indexed. For each message, relevant passages are retrieved and sent to the model as context. This technique, known as retrieval-augmented generation (RAG), is intended to ground responses in the selected references.

Before generation, each message goes through a risk check. When this step identifies signs of violence or a threat to life, the workflow stops the model request and displays support and emergency channels, such as Brazil’s 180 helpline and 190 emergency number.

My role

I am responsible for the backend and app architecture, data modeling and content indexing process. I also developed the AI service integration, risk checks, usage controls and interface.

Key features

• Situation and conversation records
• Guidance generated with context retrieved from the content library
• Recommendations and saved items
• Risk checks with support channel information
• Usage limits by period

Architecture and technical decisions

The backend uses Next.js and Supabase, with PostgreSQL, authentication and Row Level Security policies. The app uses React Native, Expo Router, NativeWind and TanStack Query.

The data model separates user and relationship profiles, conversations, messages, sources, content chunks, recommendations and saved items.

Risk checks run before the model is called. This applies the workflow directly in the application rather than relying solely on prompt instructions.

Inputs are validated and have size limits. AI usage is recorded per user and per day, and limits are checked before each request to help control costs and excessive use.

Technical challenges

• Retrieving relevant context. Chunk size needs to balance context preservation, relevance to the question and request cost.
• Separating risk checks from generation. These cases must be handled before the model produces a response.
• Controlling AI consumption. Recording usage in the database and checking per-account limits applies product rules before new requests are made.

Outcome

In development, with the landing page and main workflow available online.

Technologies

Next.js, React Native, Expo, TypeScript, Supabase, PostgreSQL, OpenAI and NativeWind.