Building Compresh (compre.sh) — context compression for LLM apps. It runs as a
drop-in proxy that replaces older conversation turns with a query-aware slice, so
long sessions stop re-sending the whole history every turn.I care about the honest version of the tradeoff: open benchmarks, and I publish
where compression loses to a raw/RAG baseline, not just the wins.
Code: github.com/compresh/compresh-benchmarks
If I bring up Compresh in a thread, that's the disclosure — I work on it.
Contact: hello@compre.sh