DevlinOS

DevlinOS is the AI operating system that runs the operational side of my businesses. It is internal software built by a team of two. I built the groundwork and the early automations, then I taught Nathan Wolfe, our operations manager, how to leverage my AI development workflow. We have built it together ever since. Everything below is current as of August 2026. This system is getting smarter and more capable every single day.

It started with financial stress

The first domain was finance, because that is where it hurt. We run a licensed career school with real contracts and real payment plans, and we did not have a dedicated finance professional or the budget to hire one. Nathan, our operations expert, was getting bogged down managing contracts, setting up ACH payments, chasing missed ones, and tracking when a student had missed enough payments that our own contract required action. Every hour of that was an hour of his attention gone from the work only he can do.

So instead of hiring a finance team we could not afford, I started building the finance domain: bank transactions sync on their own, payments match themselves to student payment schedules, and a missed payment surfaces the same day it happens instead of being discovered months later. That change alone has recovered revenue that used to slip through, and it proved the pattern the rest of the system now follows. Find the work that eats our attention, build the system that automates it, and keep our attention on the decisions that matter.

The four-and-a-half-hour meeting

The turning point for this project was a 4+ hour meeting where I taught Nathan how to get set up with my Claude Code development workflow.

Before that, we worked the way most teams do with AI: Nathan knew the operations inside and out, I would interview him like a subject-matter expert, and then I would go build automations to free up his time. Our meeting collapsed that feedback loop. Nathan hit the ground running on top of the existing system, and now he is both the expert and the implementer: he builds in the finance and admissions domains where he lives every day, I build in the content and marketing domains where I live, and we run our ideas past each other as we go.

For example, here's how the Admissions domain works now. When someone applies to the school, the system enriches the application with everything we already know, drafts a personalized email, and queues everything for review. A person approves, edits, or rejects; the system then handles the enrollment agreement, payment setup, scheduling, and course access. The judgment stays human, but there's no more busywork.

How I use it every day

My side is content. DevlinOS watches this site's search and AI-visibility performance, identifies which articles need updating and where the content gaps are, and then does the heavy lifting on drafts: it researches the topic, draws on my founder experience bank of thousands of first-hand entries from a decade of videos, posts, and braindumps, and asks me questions when the bank has a gap rather than inventing an anecdote. Then it shows me a draft, and I review and revise it alongside the AI before anything ships.

The difference is not subtle. Updating an article used to mean writing from scratch: five to fifteen hours depending on the piece, three to five days of calendar time each. With this system I recently refreshed 20 articles in under two weeks, spending about 20 minutes of my own time per article on review and revision. The same domain helps me draft LinkedIn posts, and the system triages our inboxes so that what reaches us via email is what actually needs a person.

AI drafts, humans approve

Here's the rule that's central to this system: every email, article, post, and payment action that leaves the building passes a human gate first. Anything waiting on a person lands in a single My Work queue that aggregates every draft, review, and approval across every domain, and the AI staff in Slack (Mark on finance, Eliza on admissions, Jules on marketing) can each reach only the tools their role allows, the way you would scope permissions for a human hire.

The school side handles real student records, so the system is engineered accordingly: agents work with anonymous IDs instead of names, student Social Security numbers sit entirely outside the AI's reach, every AI call is priced into a cost ledger so we know what any article or application review actually cost, and the whole thing ships like production software, with thousands of automated tests, a staging environment, and review on every change. Treating business operations as a real codebase is what makes it dependable enough to run finances and a licensed school.

Replacing the software stack

The other goal is bigger than saving time: we are building DevlinOS to replace the subscription software the businesses run on, from our email service provider on down, with systems we own outright. When the replacement effort is complete, it will save us more than $17,000 every year. We are not there yet, but we are on the path, and every replaced product is one less tool that owns our data or slows down our workflow.

Results

The practical result is capacity a small team should not have. The menial work still happens; it just happens without us, and what reaches us is the part that actually needs a person.

Where this goes

DevlinOS stays internal for now. Business owner friends keep telling me how much they want something like it, and maybe someday I'll open it up. For now, we'll keep building this system to free us up to do what we do best: interacting with students, building devlin.ai, and sharpening our AI skills as we go.

If you are an instructional designer wondering whether the AI-first way of working I write about is real, this page is the evidence: automate the work that does not need you, keep a human on every decision that does, and spend the reclaimed hours on the work only you can do. That idea runs through how I think about AI in instructional design, the curriculum at Peck Academy, and everything else on this site.