Enterprise AI usually breaks after the demo, when the system meets real data, permissions, workflows, evals, security, cost, and ownership.
Writing
Notes on AI systems, LLM security, agents, and working with models.
Before building another AI pilot, map the process, data, risk, operating model, and ROI path. This is the audit I run before recommending a build.
Enterprise agents fail when teams treat context as a long prompt. Production systems need scoped memory, retrieval, permissions, tool context, evals, and observability.
Production agents need threat modeling, permissions, evals, tracing, guardrails, and human review. Security starts where demos usually stop.
A production LLM system needs regression tests for answers, retrieval, tools, refusals, cost, latency, and human handoff. Accuracy alone is not enough.
When an LLM can call tools, security is no longer only about text. You need permissions, confirmations, audit logs, tool schemas, sandboxing, and prompt-injection resistance.
A practical note on MCP tooling: enabling many tools at once can contaminate the context window and make agent behavior harder to control.
A field note on running local models with MCP tools: where the experience is strong, what breaks, and why local AI workflows still need architecture.