Lexi AI 🤖

Project Status: (Conceptual Prototype - Potenc.IA Course | Creators of the Future with AI)

A prototype of an AI assistant focused on ethical and hyper-personalized recommendations, built with Streamlit and the Google Gemini API.

🎯 The Problem

In the information age, we suffer from the "paradox of choice" (or decision fatigue). Current recommendation algorithms are superficial and fail in two main points:

1. They Ignore Negative Filters: They don't understand niche preferences or specific moral/ethical restrictions (e.g., "I want romance series, but that don't have graphic violence scenes").
2. They Generate Dependency: Many AI systems are designed to maximize "engagement", encouraging emotional dependence and excessive use, instead of focusing on efficiency.

✨ The Solution: The "Virtual Persona"

Lexi AI solves this through a two-layer "Virtual Persona" system:

1. The "Mother-Persona": It is the AI's main SYSTEM_PROMPT. It contains the Ethical Guardrails (the "don't be a therapist" rule), the AI's personality (helpful, but not intimate), and the logic of how it should build the user's persona.
2. The "User Persona": It is the profile that the user builds interactively. It stores not only tastes (I like K-pop and cottage core), but mainly their filters and boundaries (I hate horror movies, I love series about chicken's feet).

The goal is to be an efficiency tool: a "second brain" that saves the user's research time, so they can use their rest time to rest.

🚀 Main Features (of the Prototype)

• Chat Interface: A clean and reactive interface built with Streamlit, featuring a modern, dark-themed UI.
• Dynamic Profiling (Real-Time): Users can define their "Likes" and "Negative Filters" via the sidebar. The "Update Profile" button instantly injects these preferences into the AI's context without losing the ongoing conversation history.
• Privacy-First Architecture: The project supports a "Bring Your Own Key" model. API keys entered in the UI are processed in temporary RAM only and are never saved to a database, ensuring user security.
• Ethical Guardrails: The AI is instructed (via the "Mother-Persona") to identify and redirect conversations that go out of the scope of recommendations (like requests for therapy), aiming for the user's mental health.

🧠 Key Concepts (Future Architecture)

This prototype proves "Phase 0", but the project's complete design (discussed in the conception) envisions a more robust system:

• The "Invisible Robot" (Backend): An asynchronous process that would perform data collection (Data Scraping) on social media (e.g., "fan edits" on TikTok) and forums (e.g., MyDramaList).
• Topic Discovery (Topic Modeling): Instead of using predefined tags, the "Robot" would use AI to discover relevant tags (like "funny" or "cute") by analyzing word frequency in fan discussions, allowing for niche recommendations.

🛠️ Technologies Used

• Python
• Streamlit (For the web interface)
• Google Gemini API (For the AI's brain)
• GitHub Codespaces (As a cloud development environment)

🤖 Development Methodology & Transparency

This project adopts an AI-First approach, where I acted as the Technical Product Manager and Prompt Engineer.

• Core Logic (Human-in-the-Loop): The Python code was generated with the assistance of Google Gemini, guided by my specific architectural requirements (the "Mother-Persona" system). I was responsible for supervising, reviewing, and iterating on the code to ensure the logic met the project's ethical goals.
• Frontend (Manual Effort): The CSS styling was manually written by me to achieve a custom look and dark theme (currently a work in progress), demonstrating a desire to customize the interface beyond standard Streamlit defaults.
• Educational Goal: The primary focus of this prototype was to exercise Product Conception and Prompt Engineering, leveraging AI as a tool to accelerate the coding phase while maintaining strict human control over the final output.

🏃‍♀️ How to Run the Prototype

This project was developed to run easily on GitHub Codespaces or locally.

1. Installation

• Clone or open this repository.
• Install the dependencies:

  pip install -r requirements.txt

2. API Configuration (Choose one method)

• Method A (Easy - UI): simply run the app. You will be able to enter your Google Gemini API Key securely in the sidebar.
• Method B (Developer): Create a folder named .streamlit and a file inside it named secrets.toml. Paste your key there: GEMINI_API_KEY = "YOUR_KEY_HERE".

3. Run the App

In the terminal, run:

streamlit run app.py

Streamlit will notify you to open the application in a new browser tab.

📆 Next Steps (Future Phases)

• [Phase 1 - Media]: Expand the prototype to connect to real databases (like Common Sense Media) to validate the filters.

• [Phase 2 - Style and Shopping]: Implement Computer Vision (CV) so the AI can analyze photos of clothes and recommend outfits based on styles (Cottage Core, Y2K).

• [Phase 3 - The Robot]: Build the "Invisible Robot" (backend worker) to perform data collection and topic discovery in real-time.