// HOW I BUILD
An AI-assisted skincare analysis and recommendation PWA that combines ingredient analysis, skin profiles, BPOM verification, product discovery, and recommendation workflows.
#01 / CASE FILE
A concise view of the problem, the working context, and the response behind this project.
Built as a three-person team project for Garuda Hacks, SkinSafe AI is a consumer health-literacy prototype designed to help users understand skincare ingredients, product compatibility, regulatory verification, and recommendations through a web PWA.
Skincare users often have difficulty interpreting ingredient lists, understanding compatibility with their skin profile, verifying product registration, and comparing products. The project explored how structured product and ingredient data could be combined with AI-assisted recommendations without presenting uncertain verification results as facts.
Separated the public Next.js application from the internal FastAPI AI service through a typed service boundary. Combined profile information, ingredient signals, BPOM verification, product search, and recommendation outputs into guided user flows. Designed graceful fallback behavior when external verification could not provide a reliable answer.
Used Next.js and TypeScript for the PWA and Backend-for-Frontend layer, integrating with a FastAPI AI service through typed JSON contracts, server-side service authentication, Zod validation, and structured error handling. Built live BPOM verification with normalization, caching, proxy/direct fallback, and graceful failure behavior. Developed an INCIDecoder-based ingredient signal layer indexing 4,000+ ingredient names. Implemented reusable product search, TanStack Query integration, adaptive profile and retest flows, recommendation refresh, and ingredient education interactions.
#03 / OUTCOME
BUSINESS / PRODUCT VIEW
Delivered a functional hackathon prototype integrating AI-assisted recommendations with structured product, ingredient, profile, and regulatory signals. The team recommendation pipeline was evaluated using weak-label validation with macro F1 scores of 0.9082 on validation data and 0.8217 on an unseen-brand holdout.
Open to full-time engineering opportunities and end-to-end software projects.