Retrieval systems
Fast, precise, workload-aware retrieval that does not weaken correctness.
Problems that we care about
Context infrastructure
Freshness, routing, compression, reuse, and proof for long-context systems without context rot.
Memory architectures
Temporal coherence and pollution control without becoming a generic remember-everything product.
Protocols for intelligence
Interfaces that let models, agents, and stores exchange state reliably.
Published Research articles and notes
Public notes on the technical boundaries behind freshness-aware retrieval, safe reuse, and context proof systems.
Research 001
Hybrid Semantic Cache: cache invalidation for high-dimensional similarity search.
A note on why similarity caches need invalidation semantics, freshness ownership, and traceable reuse boundaries before they can be trusted in production retrieval.
Research 002
KyroBench: context correctness under changing knowledge.
A benchmark for testing whether context systems stay current, scoped, pollution-resistant, and provable before agents act on retrieved knowledge.