The AI Brain Behind Ingredient Transparency
TL;DR:
We’re continuing to build an AI-powered ingredient interpreter for skincare, designed to help users decode formulations with scientific precision and user-first language. This update reveals the latest design evolutions, schema decisions, and GenAI behavior prototypes.
What’s New
Structured a skincare AI assistant using tagged ingredient data and strict, no-guess prompt rules.
Tested with 50 real-world queries to validate accuracy, logic handling, and tone compliance.
Achieved 86% accuracy and 96% strict compliance, with clear fallback behavior on missing or risky data.
Failures exposed edge cases like missing fields, unclear logic filters, and tone drift—now feeding next dataset upgrade.
Defining the Output Format
To make the system usable for both consumers and professionals, we defined a structured schema for ingredient insights that balances technical depth with readability.
MVP Architecture: Keon’s Local Prototype
Keon spun up a local dev stack using Ollama to prototype and benchmark LLM output across different prompt styles. We’re using:
Frontend: Lightweight React app for rapid input + result loop.
API Layer: Gin-powered REST API connecting to Ollama.
LLM Runtime: Ollama serving locally fine-tuned models for privacy and iteration speed.
Example Output
React Frontend → Gin REST API → Ollama LLM Engine
Why It Works: Fast iterations, low-latency prototyping, and versionable LLM behavior that adapts to both consumer questions and professional use cases.
Sourcing Data
While Keon stress-tests prompt consistency, I’m curating a vetted data pipeline, focusing on medically reviewed, user-readable sources. Our ingestion process prioritizes:
Evidence-based summaries
INCI-standard nomenclature
Contextual safety notes (e.g., photo-sensitivity, pregnancy concerns)
The long-term vision: ingredient insights with traceability, explainability, and trust flags baked in.
Here’s a screenshot of the early frontend where it all comes together.
Phases:
Phase 1: Testing + Iteration (Current Phase)
Integrate with LLM or retrieval-augmented model
Conduct usability testing on language clarity and risk phrasing
Align outputs with accessibility and enterprise design systems
Phase 2: Testing + Iteration
Begin user testing with real ingredient lists
Test upload and paste workflows
Build a trust-feedback loop to guide future improvements
Future Plans
Layer in user profiles (e.g., acne-prone, sensitive skin, rosacea)
Enable “smart” questions (e.g., “Is this pregnancy safe?”)
Develop a Chrome extension for ingredient popovers while browsing.