Notes from the Messy Middle
Long essays, quick lessons, and occasional scars from building AI systems that have to work outside the demo.
Agents With Brakes: Production Agents Need Control Systems
A detailed write-up on Agents With Brakes, a reference architecture for enterprise agent platforms built around policy, identity, tool access, evals, observability, approvals, cost control, memory, and failure management.
Read the essay- Question Start with the real constraint.
- Architecture Map the system around the model.
- Failure test Find where the demo gets expensive.
- Practical lesson Keep what survives contact.
Long-form thinking
Deep writeups on the systems around the model.
GuardianLLM: Child-Safe AI Should Be More Than One Moderation Checkbox
GuardianLLM is my open-source experiment in age-aware AI safety middleware for child-facing LLM applications.
PromptShield: Because Your API Key Does Not Need a Wellness Journey
PromptShield is my open-source, local-first privacy firewall for generative AI prompts.
Building KnowledgeOS: Why Enterprise AI Needs More Than RAG
KnowledgeOS is my open-source lab for the architecture enterprise AI needs after the chatbot demo: meaning, evidence, policy, evaluation, and traces.
What Computer-Use Agents Need Before We Trust Them
AgentDesk is my exploration of the policy, observability, evaluation, and cost-control layers computer-use agents need before they belong in real workflows.
AgentOps Simulator: What Happens When Your AI Workflow Starts Burning Tokens?
A small simulator for understanding agent workflow cost, routing, reliability, guardrails, and observability tradeoffs before production teaches the lesson with a bill.
Secure Tool Gateway: Agents Should Not Get API Access Just Because They Asked Nicely
A policy-driven gateway for secure AI agent tool execution with scopes, approvals, sandboxing, risk checks, budgets, and audit logs.
AI Failure Atlas: A Field Guide for How Agentic Systems Break
A field guide to prompt injection, tool overreach, infinite loops, context poisoning, and hallucinated tool calls in agentic AI systems.
Multi-Agent Planning Without the Group Chat Chaos
A small Python project showing how planner, executor, reviewer, and orchestrator roles make multi-agent workflows easier to understand.
Your Agents Need Business Cards Before They Can Work Together
A beginner-friendly A2A project showing how specialized agents discover each other, exchange tasks, and collaborate through protocol boundaries.
Your AI Assistant Should Not Get Your Whole Hard Drive
A beginner-friendly MCP project showing how AI assistants can safely read, search, and update local notes through controlled tool access.