Responsible AI (Agent / LLM Development and Usage)
Bias and fairness - LLMs learn b. from large [proprietary] training corpora & start to reproduce / amplify them.
Transparency and accountability - Due to LLMs’ incredible capabilities, disclose AI use and ensure human oversight, esp. in high-stakes domains like healthcare - directly impact human well-being (AI in latter can be treated as medical devices and, thus, regulated).
Generating harmful content - LLM may hallucinate, be prompted to generate fake/misleading info or jailbreaked to follow harmful instructions.
Intellectual property - if proprietary training data, it’s unclear when copyrighted material is used by LLM. Is AI output user’s IP, or LLM creator’s or training text author’s?
Regulation of commercial applications - due to enormous impact of AI (European AI Act).
Responsible AI - Agent / LLM Guardrails (to summarize)
NeMo Guardrails - programmable proxy layer that intercepts & modifies traffic - actively steer convos & block unsafe content in real-time at infrastructure layer.
LangSmith - real-time monitor, trace, and rigorously test whether agents are behaving correctly and safely across complex workflows. Many teams use both in tandem.
NVIDIA NeMo Guardrails
NeMo Guardrails is an open-source toolkit that sits between your application and the LLM. It actively blocks or alters inputs and outputs in real time based on programmable policies defined in its custom modeling language, Colang.
It categorizes its guardrails into five distinct stages:
Input Rails: Applied to the user's prompt before it reaches the model. These can reject the input entirely or alter it (e.g., masking Personally Identifiable Information (PII), detecting jailbreak attempts, or rephrasing for clarity).
Dialog Rails: Control the conversational flow. These rails ensure the agent stays on topic, dictate when to trigger specific tools, or force the agent to use predefined, hardcoded responses for specific intents rather than generating a new response.
Retrieval Rails: Used specifically for Retrieval-Augmented Generation (RAG). These rails evaluate the chunks of data retrieved from a database before they are sent to the LLM, allowing you to reject irrelevant chunks or mask sensitive enterprise data.
Execution Rails: Applied to the inputs and outputs of custom actions or external tools. They validate that the LLM is calling a tool securely and that the tool's output is safe to pass back into the system.
Output Rails: Applied to the final generated response before it reaches the user. These rails can reject or edit the output to ensure content moderation, detect hallucinations (fact-checking against the retrieved context), and filter out toxicity or sensitive data.
LangSmith (LangChain)
Not for interception, but is an observability and evaluation platform via continuous evaluators running on application's traces—offline during dev or online in real-time production.
Heuristic / Deterministic Checks: Hard-coded programmatic validations. These act as strict guardrails to ensure the agent's output adheres to a specific JSON schema, compiles correctly (if generating code), or passes regex filters (like strict PII detection).
LLM-as-a-Judge Evaluators: Semantic graders that score outputs against custom criteria. You can set up evaluators to detect hallucinations, measure groundedness (how well the answer aligns with the source context), check for toxicity, or ensure conciseness.
Multi-Turn Trajectory Guardrails: Because LangSmith traces the entire "internal monologue" of an agent, you can evaluate the agent's step-by-step reasoning. You can check if a safety policy was triggered at the right stage (e.g., "Did the agent refuse the request when required?" or "Did it avoid revealing its system prompt during the interaction?").
Online Evaluations: You can sample live production traffic and run your evaluation guardrails in real time. This acts as an early warning system to detect model drift, unexpected tool usage, or quality degradation in the wild, triggering alerts if safety scores drop below a set threshold.
References
Encyclopedia of LLMs and Foundational Models - it has all the GenAI concepts described in a concise manner, can be useful to search for specific definition, no guarantee it was not generated by LLMs (downloaded to Books and MLBasics too)
TranslateGemma - new SOTA that runs on edge and laptops (Jan 2026).
Ollama runs locally (including agents)
Top AI GitHub repositories: https://blog.bytebytego.com/p/top-ai-github-repositories-in-2026?utm_source=substack&publication_id=817132&post_id=189059613&utm_medium=email&utm_content=share&utm_campaign=email-share&triggerShare=true&isFreemail=true&r=wakvt&triedRedirect=true