Filed under
AI
39 entries
Security·7 min read·
A Correct Answer Doesn't Mean the Agent's Memory Is Clean
A synthetic memory-recovery drill shows why a correct task result and a clean derived state are separate checks.
Security·6 min read·
The Database Gets a Vote on Your Read-Only Agent
A disposable PostgreSQL lab tests what an agent-facing database role can actually do, including the SELECT that writes through a privileged function.
Security·8 min read·
The Policy Is the Part That Moves
A proved gate is only as good as the policy it enforces, and the policy is the part that changes while the agent is running. What OSCAL and the agent-authorization drafts both assume, and what breaks.
Systems·4 min read·
The Filesystem Should Start as a Specification
SYSSPEC treats a filesystem specification as the thing an agent edits, then asks generated code to live up to the contract.
Security·7 min read·
You Can't Prove the Agent. Prove the Gate.
Formal verification for AI-agent security. You can't prove a probabilistic model does the right thing — but the deterministic gate in front of it is small enough to prove outright. Dafny proofs, a Rego twin, and differential testing, following the method AWS used for Cedar.
Security·7 min read·
The Firewall Stays Put. The Agent Improvises.
Rewriting security controls for probabilistic agents. Static control catalogs assume systems that do what they're told; AI agents don't — so here's how to express agent guardrails (tool allowlists, secret-egress denial, audit) as machine-checkable OSCAL component definitions.
Security·13 min read·
The Sandbox Isolates the Agent. It Doesn't Isolate the Secret.
MicroVMs and namespace sandboxes give AI coding agents mature filesystem and process isolation. Most open-source ones still hand credentials to the agent directly. Here's the proxy-injection pattern that fixes that, and what it doesn't fix.
Typescript·7 min read·
live-coding-music-mcp: Handing Claude a Browser and a Drum Machine
An MCP server that lets Claude live-code music in Strudel through a real, visible Chromium browser — no mocked audio API. What it does, what broke while I built it, and the engineering that came out of my first real browser-automation project.
AI·10 min read·
A Month of Modularization: nexus-agents in April 2026
Extracting benchmarks into a standalone package was the punchline. The setup was a month of governance, skills, security, and pipeline discipline that made the extraction possible in an afternoon.
Security·6 min read·
Investigating the Trivy Supply Chain Compromise with AI Agents
How I used AI-assisted investigation to triage the trivy-action supply chain attack across my homelab repos — and some thoughts on weekend incident response and community notification gaps.
AI·12 min read·
Building Nexus-Agents: What I Learned Creating a Multi-Model AI Orchestration System
The engineering story behind nexus-agents, a research-backed multi-model orchestration system that coordinates Claude, Gemini, and Codex through consensus voting, adaptive routing, and graph workflows.
AI·8 min read·
Consensus Voting With AI Models: When Three Opinions Beat One
How multi-model consensus voting catches blind spots that single models miss. The research behind adversarial roles, Bayesian aggregation, and structured deliberation across Claude, Gemini, and Codex.
AI·8 min read·
From RouteLLM to Contextual Bandits: How Research Papers Shaped My Model Router
How I went from naive round-robin model selection to a five-stage routing pipeline backed by RouteLLM, TOPSIS, and LinUCB research. The failures that led to each improvement.
LLM·10 min read·
LLM-Powered Security Alert Triage with Local Models
Automate security alert analysis using local LLMs (Ollama) for privacy-preserving incident response. Reduce alert fatigue with AI-powered triage without cloud dependencies.
AI·16 min read·
From 150K to 2K Tokens: How Progressive Context Loading Revolutionizes LLM Development Workflows
Optimize LLM workflows with progressive context loading—achieve 98% token reduction using modular architecture for efficient production deployments.
AI·9 min read·
From Claude in Your Terminal to Robots in Your Workshop: The Embodied AI Revolution
Deploy Vision-Language-Action models for embodied AI robots—integrate physical world interaction with security considerations for homelab automation.
AI·11 min read·
AI as Cognitive Infrastructure: The Invisible Architecture Reshaping Human Thought
Understand AI cognitive infrastructure shaping how billions think—explore societal effects of language models transforming from tools to thought systems.
AI·7 min read·
Supercharging Development with Claude-Flow: AI Swarm Intelligence for Modern Engineering
Claude-Flow orchestrates multiple AI agents against one task. What it does, what its published numbers actually are, and why the CLI in most write-ups never existed.
AI·14 min read·
Down the MCP Rabbit Hole: Building a Standards Server
Build MCP standards server for Claude AI—implement Model Context Protocol for intelligent code standards and context-aware workflows.
AI·8 min read·
Exploring Claude CLI Context and Compliance with My Standards Repository
A CLAUDE.md is a prompt, not a loader. What a standards repository can and cannot enforce, and where the real enforcement has to live.
AI·8 min read·
Local LLM Deployment: Privacy-First Approach
Deploy local LLMs for privacy-first AI—run language models on homelab hardware with model selection, optimization, and deployment strategies.
AI·7 min read·
Teaching AI Agents to Ask for Help: A Breakthrough in Human-Robot Interaction
A robot given an ambiguous instruction usually guesses. New work trains agents to ask a clarifying question instead, using LLM-generated rewards.
AI·17 min read·
Fine-Tuning LLMs in the Homelab: A Practical Guide
Fine-tune LLMs on homelab hardware with QLoRA and 4-bit quantization. Train Llama 3 8B models on RTX 3090 with dataset prep and optimization strategies.
AI·6 min read·
Securing Your Personal AI/ML Experiments: A Practical Guide
Secure personal AI experiments with model isolation and network segmentation—protect LLM deployments using privacy controls and threat modeling.
AI·14 min read·
Context Windows in Large Language Models: The Memory That Shapes AI
What context windows actually cost in memory and latency, why capacity and recall are different things, and where retrieval beat a bigger window in my homelab.
AI·21 min read·
AI: The New Frontier in Cybersecurity – Opportunities and Ethical Dilemmas
Deploy AI-powered cybersecurity with automated threat detection—reaching 73% precision on anomaly detection catching attacks my signature-based IDS missed.
AI·14 min read·
Large Language Models for Smart Contract Security: Promise and Limitations
Testing GPT-4, Claude and Llama 3 against 47 smart contracts—80% reentrancy detection, but 60% of everything they flagged was wrong, which is substantially noisier than the static analyzer they were supposed to replace.
AI·16 min read·
GPU Power Monitoring in My Homelab: When Machine Learning Met My Electricity Bill
What three months of instrumenting an RTX 3090 actually showed about the cost of running LLMs at home—including the optimizations that worked and the one that didn't.
AI·15 min read·
AI Meets Edge Computing: Transforming Real-Time Intelligence
Deploy AI edge computing with YOLOv8 and TensorFlow Lite—achieve 15ms latency for real-time inference on Raspberry Pi with local processing for privacy.
AI·12 min read·
Learning from Nature: How Biomimetic Robotics is Revolutionizing Engineering
Design biomimetic robots inspired by nature—implement gecko adhesion, swarm intelligence, and soft robotics using billions of years of evolution.
AI·18 min read·
The Evolution of High-Performance Computing: Key Trends and Innovations
Deploy high-performance computing with parallel processing and distributed systems—access supercomputer capabilities through cloud HPC for AI workloads.
AI·15 min read·
Multimodal Foundation Models: Capabilities, Challenges, and Applications
Build multimodal AI systems with GPT-4 Vision and CLIP—process text, images, and audio together for next-generation foundation model applications.
AI·45 min read·
AI Learning in Resource-Constrained Environments
Train AI models on resource-constrained hardware with quantization, pruning, and distillation—run GPT-3 capabilities 100x faster through compression.
AI·11 min read·
Mastering Prompt Engineering: Unlocking the Full Potential of LLMs
Master prompt engineering with few-shot learning and chain-of-thought techniques—improve LLM response quality by 40% through systematic optimization.
AI·11 min read·
The Ethics of Large Language Models
Address LLM ethics including bias, privacy, and accountability—implement responsible AI frameworks for large language model deployment in production.
AI·16 min read·
Retrieval Augmented Generation (RAG): Enhancing LLMs with External Knowledge
Build RAG systems with vector databases and semantic search—eliminate LLM hallucinations and ground responses in verified knowledge for trustworthy AI.
AI·15 min read·
The Transformer Architecture: A Deep Dive
Master transformer architecture with self-attention and positional encoding—understand the foundation of GPT-4, BERT, and modern language models.
AI·13 min read·
Open-Source vs. Proprietary LLMs: A Battle of Accessibility, Customization, and Community
Compare open-source vs proprietary LLMs with Llama 3 and GPT-4 benchmarks—understand performance, cost, and customization trade-offs for production.
AI·8 min read·
The Deepfake Dilemma: Navigating the Threat of AI-Generated Deception
Deepfake detection works well on data it has seen and falls apart on data it hasn't. Why the fix is partly cultural, not just technical.