Starting from experiments,
wiring models into real engineering.
Notes on AI applications, agents, RAG, MCP and evaluation practice.
Fashion Item Matcher
A same-item retrieval system for fashion, built for content placement. I owned the vector index layer, the Kafka CDC ingestion pipeline, price filtering and AI cost governance. Built on FastAPI, Elasticsearch nested dense_vector and multimodal models, it serves image-to-item, text-to-item and content-style matching over 120k+ products and 210k+ SKU vectors.
Ying+ · AI Investment Review & Position Advice
The AI content engineering practice inside the Ying+ multi-client investment app: market, sector, holdings and watchlist data are turned into a structured Context, and DeepSeek generates an AI morning brief, daily review and position advice, with data-readiness checks, idempotent generation, back-office review and a permission/billing flow. Productization has stopped; it is kept as a record of AI application engineering and financial-data systems.