Agents, retrieval systems, and the infrastructure that holds them together. Some shipped to production, some still on the workbench.
Multi-agent orchestrator that coordinates task delegation across LLM backends. Agents negotiate priorities through a consensus protocol — no single point of failure, no bottleneck coordinator.
Research agent that reads academic papers, extracts claims, and builds connected knowledge graphs. Uses recursive reflection to self-correct before surfacing results.
RAG engine built for large codebases. Hybrid search (dense + sparse vectors) with semantic reranking. Sub-100ms queries over gigabytes of indexed code. Written in Rust because Python couldn't keep up.
PDF extraction pipeline that actually understands tables. Spatial-aware parsing handles multi-column layouts, merged cells, and nested headers that break every other tool.