Hello.

I build search and retrieval systems and then run them myself, on six machines scattered around the house and mostly piled up near the router. When something on this site is slow, I know which box to go and look at.

That is most of why I do it this way. If I own the ingestion, the model, the GPU and the network in between, every problem is mine to find, and nothing important sits behind an API I cannot see inside. Globirie has taken in over 800,000 articles on that setup, and the retrieval backend underneath it keeps two models on one GPU without either one starving the other, which took much longer to get right than I expected. Before those there was an opinion search engine over 19,187 forum posts and articles, where most of the real work turned out to be cleaning the corpus.

I study Computer Engineering at NTU, with a second major in Data Analytics. Most of what I know about running systems came from running them badly first.