Lakshya Sharma

You can build hybrid retrieval pipelines (BM25 plus dense vectors with reranking), orchestrate recursive long-context reasoning over oversized corpora, manage four-store typed memory with a promotion gate, verify answers against retrieved evidence, and safely execute model-generated code in a sandboxed kernel.

Engineers implementing or debugging retrieval-augmented generation systems, long-context reasoning engines, persistent memory stores, or evidence-verification gates. They suit developers deciding between hybrid retrieval and recursive reasoning paths, or hardening sandboxed execution of untrusted model-generated code.

Use rag-retrieval for standard document or codebase question-answering via the retrieve() interface. Escalate to rlm-orchestrator only when material exceeds any context window, such as full-repo searches or literature-review aggregation. The refinement skill governs retry-versus-escalate decisions between them.

It implements a shared evidence-sufficiency gate that scores whether each answer is backed by retrieved evidence, computed values, or explicit inference before returning. It also provides methodology for fact-checking third-party claims, benchmarks, and documented features against primary sources before committing them.

The code-execution skill defines process isolation tiers, filesystem and network restrictions, and REPL output truncation within the RLM kernel sandbox. It is a policy and checklist skill guiding tier selection for local development, untrusted input, or production deployments, governing the rlm-orchestrator kernel.