Black LLAB open-sources dynamic LLM routing, agent sandboxes

A solo mechanical-engineer-turned-dev independently reconstructing frontier lab infrastructure is exactly the kind of project the LocalLLaMA community exists for — and the architecture choices here are genuinely interesting, not just vibe-coded boilerplate.

• –The complexity grader (Mistral 3B scoring prompts 1–100) is a pragmatic alternative to expensive classifier models; routes cheap queries local and reserves Opus/Perplexity for heavy lifting
• –Docker sandboxing via OpenClaw means agents can write files, execute code, and scrape the web without touching the host OS — a real safety gap in most homebrew agent setups
• –Hybrid RAG stack goes deeper than most: HyDE, reciprocal rank fusion, cross-encoder re-ranking, and a NetworkX knowledge graph all layered together
• –XML-boundary context isolation and rolling summarization show awareness of "Lost in the Middle" degradation — a detail most hobbyist RAG projects skip
• –Score 0, zero comments on Reddit at time of scrape — this is brand new; the architecture deserves more eyeballs than it's gotten