January 1, 2026
Perpetual Insight
I've had a deep passion for luxury watches for as long as I can remember. There's something about the craftsmanship, the heritage, and the sheer mechanical artistry of a fine timepiece that I find endlessly fascinating — and I know I'm far from alone in that. The luxury watch community is massive and passionate, but navigating it as a buyer can be overwhelming. Which reference fits your wrist size? Which model aligns with your lifestyle? Which pieces hold value? How long is the waitlist?
As someone equally passionate about machine learning, I saw a natural opportunity to combine both worlds. Perpetual Insight is the result — a luxury watch AI assistant designed to help guide enthusiasts toward the watch that truly fits their needs, preferences, and lifestyle.
Perpetual Insight is a personalized recommendation engine built for the luxury watch market. It combines Retrieval-Augmented Generation (RAG) with structured market data to surface the most relevant Rolex watch recommendations based on user preferences and wait time signals.
The system is built around a hybrid retrieval and generation pipeline:
- Vector Retrieval: A Milvus vector database stores dense semantic embeddings of Rolex watch specifications, historical data, and market context. Queries are embedded using the BGE-m3 sentence-transformer to perform high-density semantic search.
- LLM Generation: Retrieved context is synthesized by a language model into natural language recommendations that maintain brand-aligned tone and factual accuracy.
- Personalized Recommendations: Combines semantic understanding of user queries with structured market data — including wait time signals — for highly tailored results.
- ETL Pipeline: A custom pipeline scrapes and ingests historical watch data and technical specifications into Milvus, providing the system with a rich, structured knowledge base.
- LLM-as-a-Judge Evaluation: Uses Llama 3 as both the reasoning engine and evaluator, scoring the quality of generated recommendations to ensure outputs meet high-fidelity standards for accuracy and brand-aligned tone.
- Python: Core pipeline and orchestration logic.
- Milvus: Vector database for semantic storage and retrieval.
- BGE-m3 (Sentence-Transformers): High-density semantic embeddings.
- RAG: Retrieval-Augmented Generation architecture.
- Llama 3: LLM backbone for generation, reasoning, and evaluation.
- ETL Pipelines & Scraping: For data ingestion and knowledge base hydration.
Related projects
Bookstore Application
A modular iOS bookstore application built with VIPER architecture.News App
A news aggregation iOS app featuring real-time updates and fake news detection.
