I'm currently deep in the intersection of agentic AI systems and distributed state management. Specifically, I'm exploring how multi-agent pipelines handle context retrieval at scale — where vector search recall quality (tuning HNSW parameters like efSearch and efConstruction) directly determines whether an agent reasons correctly or hallucinates a plausible-but-wrong answer. This connects my recent work on HNSW indexing and structured LLM outputs into a single production problem: building retrieval systems that are both fast and semantically precise enough to serve autonomous agents reliably.

On the distributed systems side, I'm studying how CRDT-based consistency models can be extended beyond collaborative text editors into broader application state — specifically whether the algebraic guarantees that make Y.js work for document sync can be applied to shared game state or multi-user configuration trees without a central coordination server. The goal is understanding where CRDTs break down and where you still need something closer to Raft-style consensus.