BeautyGPT

Consumers often struggle to find beauty products suited to their unique skin types and concerns, leading to frustration and decreased satisfaction. BeautyGPT addresses this by leveraging Generative AI to deliver personalized product recommendations and streamline the product discovery process.

In the early stages of the project, I explored multiple datasets to evaluate different approaches to powering BeautyGPT’s recommendation engine.

I reviewed the Sephora Product and Review Dataset (Kaggle), which offered extensive product listings and customer reviews but proved too brand-specific and inconsistent to support a generalized recommendation engine.

I also considered expanding BeautyGPT into computer vision by testing the Skin Segmentation Dataset (UCI) and the Skin Disease Dataset (Kaggle). These datasets could have enabled skin type or skin condition detection through image analysis. However, I deprioritized this approach due to its complexity, privacy concerns, and the technical requirements of building a reliable classifier without an engineering team.

Instead, I focused on the Skincare Ingredients - INCI List Dataset (Kaggle), which provided structured ingredient information, including benefits, recommended use cases, and skin type compatibility. I cleaned and enriched this dataset to map ingredients directly to common skin concerns such as dryness, sensitivity, acne, and hyperpigmentation. This dataset formed the basis of BeautyGPT’s knowledge base and recommendation logic.

The development of BeautyGPT began with user research to identify common challenges in beauty product discovery, such as compatibility with individual skin concerns and the overwhelming number of choices available.

Using the cleaned and structured skincare ingredient dataset, I built a lightweight recommendation engine leveraging OpenAI’s API and AgentHost.AI, which allowed me to deliver an interactive chatbot without the need for an engineering team. The system matched user-inputted skin concerns to ingredients and product characteristics, generating helpful and contextually relevant product suggestions. I guided this process through iterative prompt testing and refining of the knowledge base to balance generalization with helpful specificity.

To formalize the project, I created a comprehensive Product Requirements Document (PRD) outlining technical requirements, feature definitions, and user scenarios. Throughout MVP development, I conducted iterative testing and gathered user feedback to validate recommendations and improve the conversational quality of the chatbot.

BeautyGPT is now live as an embedded prototype on my portfolio, allowing users to interact with the tool and receive tailored skincare recommendations.

To power BeautyGPT’s recommendation engine, I designed a structured system of prompts to ensure responses were accurate, inclusive, and tailored to each user. The chatbot guides users through a simple conversational flow, collecting information on their skin type, age, and concerns before providing recommendations. The prompt system is also designed to be respectful, age-conscious, and focused only on beauty-related topics.

A simplified version of the instruction prompt:

[You are a skincare expert. Greet the user warmly and collect information about their skin type, age, and skin concerns. Use the uploaded knowledge base to recommend products or ingredients tailored to their needs. Avoid answering questions outside of skincare, haircare, makeup, or beauty-related diet topics. Be polite, supportive, and concise.]

The full prompt included additional guidance such as:

• Asking for clarifications when needed

• Explaining recommendations in simple, user-friendly terms

• Providing inclusive and ethnicity-aware advice

• Encouraging users to return as their skin needs change

Through iterative testing, I refined this prompt system to balance approachability and technical accuracy, ensuring BeautyGPT delivered recommendations that were relevant and actionable for users.

The global skincare market is valued at over $150 billion and continues to grow, fueled by demand for personalized, inclusive, and tech-enabled beauty experiences. BeautyGPT addresses this market opportunity by offering a scalable AI-driven solution for beauty brands, eCommerce platforms, and DTC businesses looking to enhance customer satisfaction and reduce product returns. The MVP demonstrates how an AI-powered assistant could be adapted into a commercial product through white-label integrations, affiliate models, or direct-to-consumer offerings - delivering value through both customer experience and business outcomes.

To power BeautyGPT’s recommendation engine, I designed a structured system of prompts to ensure responses were accurate, inclusive, and tailored to each user. The chatbot guides users through a simple conversational flow, collecting information on their skin type, age, and concerns before providing recommendations. The prompt system is also designed to be respectful, age-conscious, and focused only on beauty-related topics.

A simplified version of the instruction prompt:

[You are a skincare expert. Greet the user warmly and collect information about their skin type, age, and skin concerns. Use the uploaded knowledge base to recommend products or ingredients tailored to their needs. Avoid answering questions outside of skincare, haircare, makeup, or beauty-related diet topics. Be polite, supportive, and concise.]

The full prompt included additional guidance such as:

• Asking for clarifications when needed

• Explaining recommendations in simple, user-friendly terms

• Providing inclusive and ethnicity-aware advice

• Encouraging users to return as their skin needs change

Through iterative testing, I refined this prompt system to balance approachability and technical accuracy, ensuring BeautyGPT delivered recommendations that were relevant and actionable for users.

BeautyGPT is designed to demonstrate how AI-powered recommendations can improve the beauty shopping experience. 

The tool aims to:

• Increase user engagement by 25% by delivering relevant and personalized product suggestions.

• Reduce product return rates by 15% through improved product-to-user fit.

• Achieve 90% recommendation accuracy based on early user feedback.

• Improve conversion rates by 20% by helping users make faster and more confident purchase decisions.

• Boost customer satisfaction by offering tailored solutions aligned with user-specific skincare needs.

These projected outcomes are aligned with the product’s success metrics defined in the PRD and are supported by qualitative feedback from demo users.

Looking ahead, potential next steps include collaborating with data scientists to enhance the recommendation engine, scaling the dataset, and expanding functionality through potential brand or platform partnerships. BeautyGPT could also evolve into a white-labeled tool for beauty brands or eCommerce platforms seeking AI-powered personalization.

BeautyGPT demonstrates my ability to leverage Generative AI to deliver a functional and user-focused product, balancing technical feasibility and business value. The project highlights my approach to product management: building solutions that solve real user problems, even without an engineering team, while delivering measurable impact.