I’ve spent twenty years inside systems that aren’t allowed to fail.

I’m Johnmicah Potter. I run JJP as a principal-led practice: I sit with executives to decide what AI is for, own the architecture that answers it, and stay accountable through delivery until the thing is running in production and can stand up to a security review. Leading engineers is most of what I’ve done. Five architects setting the architecture and development standards a much larger engineering organization built against. Five platform engineers — their standups, their sprint planning, their retros, their one-on-ones. Front-end teams where I did the interviews, then the pairing and the code review. A startup’s entire engineering organization, built from nothing. And I still write the hardest parts myself.

The two decades before that were spent inside large enterprises, mostly healthcare and financial services, where software is regulated, audited, and load-bearing. I started as a developer. I ended up a senior IT architect, which mattered less for the title than for what came with it: the standards everyone else had to build against. Along the way I stood up a new team to support the AEM and Node.js infrastructure and staffed it myself. I built a CI/CD platform that standardized Terraform deployments to Google Cloud, and unified delivery for hundreds of applications across wildly different stacks. One platform I built was used by more than 1,000 engineers. Others served millions of people who never knew my name.

When production LLMs arrived I was already inside the systems they’d have to work in. At a Fortune-500 financial institution I co-invented a hierarchical multi-agent code-generation platform: agents fluent in the firm’s own component libraries, reading real Figma designs, writing applications and testing their own output before a human ever saw it — running on Anthropic models in AWS Bedrock, with deliberate context engineering and a live observability map leadership could watch. It went into production and it’s still there. I’m a co-inventor on a filed patent for it. Around it I built the rest of what production AI needs: an MCP server doing speaker-attributed semantic search across millions of meeting-transcript segments, multi-agent code review, agentic sprint operations, semantic codebase indexing, and the context and caching strategies that keep the bill defensible.

The other half of the job is what makes AI survivable. On a regulated healthcare cloud estate I owned the security vulnerabilities, policies, and remediations, stood up the infrastructure behind 17 production AI applications, pushed automated security scanning across a portfolio of 30, and cut cloud spend 30–50%. Since 2021 I’ve also sat as an executive advisor inside a large healthcare organization, the person leadership brings an AI proposal to before it gets funded. For a high-growth startup I built the engineering organization from zero: architecture, infrastructure, and the team. And I’ve spent years teaching engineers agentic design patterns and how to deploy this responsibly, through pairing, code review, and training programs, because a system nobody else can run isn’t finished. That’s what this practice does: take an organization from AI ambition to governed production, and leave engineers behind who can build the next one themselves.

What I believe

Anyone can get an answer out of an LLM now. The work is everything around it: grounding it in your data, proving what it did, keeping it inside policy, and running it at a cost you’d defend in a budget meeting. Trust is an engineering property. You earn it with retrieval you can trace, evaluations that run before release, logs an auditor can read, and a person in the loop where the stakes justify one. Compliance is the smaller reason for all of that. The patient, the claimant, the person waiting on a decision never chose to be in the room with your system, and they’re the ones who live with what it gets wrong. Systems built this way get adopted. The rest get quietly switched off.