FARSIX: Deterministic Cognitive Pipeline Orchestrator & Guardrails Engine

Overview

Generative AI integration into mission critical enterprise software faces a core architectural challenge: high-parameter Large Language Models are inherently probabilistic, while production execution layers require absolute determinism. Standard API wrappers and basic regular expression filters fail to catch subtle cognitive hallucinations, structural parameter errors, or logical contradictions. FARSIX addresses this challenge by establishing a dual-layer cognitive architecture that separates reasoning from output verification inside a Directed Acyclic Graph state machine. The primary processing stage delegates heavy visual and analytical tasks to cloud-hosted NVIDIA NIM microservices. Visual input ingestion and UI layout analysis are processed by Nemotron-Nano-VL-8B, while complex analytical deductions and structured operational reports are generated by Llama-3.1-70B-Instruct. To prevent malformed or hallucinated payload execution, FARSIX implements a dual-model semantic auditing pipeline powered by NVIDIA NeMo Guardrails. Rather than relying on static syntax checks, a dedicated Llama-3.1-8B-Instruct auditing model evaluates generated reports against semantic rules written in Colang. The auditor checks intent classification, logical consistency, and domain safety rules in real time. When the auditing layer detects a rule violation or logical contradiction, the orchestration engine physically blocks downstream pipeline execution. It logs the exact failure state and triggers an automated recovery loop, routing the execution context back to the primary reasoning layer with explicit failure context for correction. The system includes a real-time web dashboard that renders the state machine as a dynamic SVG graph. Developers can monitor node execution states, inspect token throughput, track validation latencies, and review semantic audit logs live. The entire orchestration framework is modular, allowing custom Colang safety policies and alternative model endpoints to be integrated without altering core pipeline logic.