Samuel Grummons

AI agents have the potential to enable powerful workflows. What distinguishes an AI agent from a language model is its ability to reason and act, or in technical parlance, ReAct. However, AI agents are only as safe as the data boundaries we enforce. Agentic systems must be built with security and privacy at the architectural level.

This article discusses both anti-patterns and best practices for building AI Agents in LangGraph with secure access to personal data.

Implemented a retrieval-augmented generation (RAG) application using Milvus for vector storage, LangChain for LLM interactions, and Streamlit for the front end. The application now uses LangGraph to implement a reasoning and acting (ReAct) framework for tool usage.

In Harry Frankfurt's book On Bullshit, Frankfurt distinguishes between lies and bullshit. The liar is concerned about the truth and ensures the other party does not find it. The bullshitter, however, is not interested in whether what they say is true or false; truth is irrelevant to their goal.

I have often joked about developing a "bullshotmeter," a tricorder-type device that would beep in the presence of bullshit. This project is an attempt to bring that tech to life.

The hypothesis is that bullshit has its own semantic patterns. Maybe it's more bombastic, more sure of itself, less nuanced, more focused on eliciting some kind of response from the other party. This project is an attempt to model those semantic patterns and use them to develop a classifier for bullshit.

This is my final assignment for the Hugging Face AI Agents Course. I scored 70% on the GAIA Level 1 Benchmark using gemini/gemini-2.5-pro-preview-03-25 as the core model. I also use smaller models to offload some of the tasks. The README includes some of my lessons learned while building the solution.

An article exploring the risks of agentic AI systems and mitigation techniques recommended for enterprise adoption of AI-driven automation. Agentic AI systems can autonomously plan and execute complex tasks using enterprise tools, offering significant benefits—like advanced IT support—but also introducing unique security risks. Unlike deterministic systems, agentic systems make non-linear decisions, creating challenges for secure tool access. Improper authentication or excessive permissions can expose enterprises to prompt injection attacks. The key mitigation is enforcing least privilege and building secure wrappers around tools with enterprise-grade authentication. As protocols like MCP evolve, vendors and enterprises must prepare for secure AI integration.