15 min read
A Structured Benchmark for AI Paper Review
Refine won 90.4% of 1,349 head-to-head matches against single-shot LLM reviewers and scaffolded review systems on 150 economics preprints.
Blog
Notes on research, products, and the work we do.
Browse by tag
13 min read
How jemalloc cut our production memory by 47%
A FastAPI service sat pinned at its 1 GiB memory cap. The cause was glibc fragmentation under bursty async load; swapping the allocator to jemalloc cut average production RSS by 47%.
