AI optimization experiments. Each experiment is self-contained with its own pyproject.toml and uv environment — cd into the directory and uv run to get going.
| Experiment | Description |
|---|---|
| prompt-compression | Local prompt compression with LLMLingua-2 (BERT-based, ~500MB). FastAPI + browser UI. |
| prompt-proxy | Transparent compression proxy for Claude Code & Codex. Rule-based structural transforms + LLMLingua-2 semantic compression with frozen/middle/hot zone architecture. |
| agent-memory | Persistent agent memory via RAG + LanceDB + sentence-transformers. Extraction-first writes, time-weighted retrieval, MMR diversity, consolidation. CLI + FastAPI UI. |