14/04/2026
We’ve all seen it: a beautiful LLM response… that quietly invents a citation, a number, or a logical step. In consumer chat it’s annoying.
In scientific pipelines, medical summarization, finance, or agent swarms — it’s catastrophic. That’s why we built Director-AI: the only production-grade, real-time hallucination guardrail that sits between any LLM backend and the user and cuts the t**en stream dead the moment coherence drops. Not post-hoc scoring. Not another slow LLM-as-judge loop. Real dual-entropy scoring: lightweight rules + embeddings + 0.4B FactCG-DeBERTa NLI + 6 advanced RAG strategies against your knowledge base.
Rust-accelerated hot paths (9.4× speedup).
SwarmGuardian for multi-agent systems (LangGraph, CrewAI, OpenAI Swarm, AutoGen).
Token-level halt + full audit trail for EU AI Act compliance. 75.6 % balanced accuracy on LLM-AggreFact at 14.6 ms latency. It was born as our internal tool at ANULUM Institute for high-stakes scientific computing. Now it’s open. AGPL for research/personal use.
Commercial dual-license available for closed-source, SaaS, or enterprise deployments. If you’re an AI engineer, founder, or team lead tired of babysitting hallucinations — check it out:
→ https://github.com/anulum/director-ai
Full docs + quickstart (6 lines): https://anulum.github.io/director-ai
Looking for serious clients ready to pilot in medical/finance/scientific stacks.
Also speaking with investors who understand the safety moat in agentic AI. DM or email [email protected] No hype. Just engineering that solves a problem most guardrails pretend to solve.
Real-time LLM hallucination guardrail — NLI + RAG fact-checking with t**en-level streaming halt
