Technical depth, in the open.

A high-recall retrieval kernel for cases where standard HNSW configurations fall short.

Characterizing where and why graph-index recall drops past 1M vectors.

Coordination patterns for multi-step agent workflows with human-in-the-loop gates.

Durable operational memory: what an agent system should remember between runs.

Using structure between documents to improve multi-hop retrieval quality.

Late-interaction and multi-vector representations under real latency budgets.