Technical Blogs

Writing about ML systems, GPU programming, inference optimization, and things I learn along the way.

---

• CacheClip: making RAG prefill 3.33× faster without giving up the answer
  29 Aug 2026 · 18 min read

  How a 135M-parameter helper model decides which KV cache entries the 14B model actually needs to recompute.

• CacheBlend: reusing KV caches for RAG without losing cross-attention
  14 Aug 2026 · 16 min read

  Recomputing about 15% of the tokens per layer buys you the speed of KV reuse and the quality of a full prefill.