What Our Interns Shipped, and What We Changed So They Could

Last week, I argued that if AI takes over the entry-level rungs while senior roles still demand entry-level experience, the senior pipeline empties out within a few decades. I also said we had just rebuilt what entry-level means at our own company. Here is how.

PlumOS hosted 2 interns this summer, Rohit and Kenley. Between them, they merged 45 pull requests, rewrote our entire data room, built part of our communication strategy, and ideated new features and fixes on our production platform that serves clients every day. Rohit is a CS student. Kenley studies Economics and Literary Arts.

They are both exceptional, and most importantly, self-motivated. Their profile is not unique, and they would have done well in any era. What changed is just how much they can contribute and how much the setup around an entry-level person now determines their learning, output, and impact.

The Old Shape of an Internship

Rohit had a different CS internship last summer, and he describes it the way most of us remember our own internships: he owned one small piece of a larger system, waited on other people for context, and rarely saw a feature all the way through.

Context lived in senior heads and was expensive to transfer, and a mistake in production was costly. His last company was large and established, so one could argue that size demands more security. But it goes further than that. Structurally, it was not feasible for a company to have someone constantly answering questions, explaining what good looks like and verifying that the work was clean and consistent. AI changes that.

AI now can be a constant companion that answers the hundreds of questions that might seem obvious to someone established but would be impossible to guess if you were a new hire or someone at the beginning of their career. While AI can be a phenomenal sonic screwdriver, whether your team will get gold or slop completely depends on access, structure, and the framework you put around it.

What We Changed

I said last week that this starts with capturing the tribal knowledge that lives only in your veterans’ heads. That sounds like a documentation project. In reality, it is about framework and process.

We decided which information the company needed to keep, built a repository for each kind, and wired the capture to run on its own. Our meeting and call transcripts get reviewed automatically, and what matters in them lands in those repositories, searchable through our internal tools. The knowledge base gets richer every week the company runs.

The code side works the same way. Our repository carries its own rules and skills, so the AI an intern pairs with already knows our conventions, and they can implement a real ticket without memorizing the whole stack.

The path to production is the same one our senior engineers use. Open a pull request, get automated review on every change, merge to dev, ship to prod; automations keep the queue in sync and merge when the checks come back green, so senior review stops being the bottleneck. Interns get real access from the start, including production logs, where Rohit learned to investigate live failures the way a staff engineer does, and customer threads where problems actually surface.

The team norm is to post what you found in Slack before you open the pull request, so discovery is visible and feedback arrives early. Daily syncs and pairing on the hard problems do the rest; when Kenley hardened our live-review infrastructure, she did it alongside Dan, our head of engineering. Because AI now handles so much of the onboarding, Dan had the time to go deep with her on that one project. Mistakes are expected, so the system is built so they get caught while they are still cheap.

AI runs through all of it in 3 roles. It writes the routine code. It shows a newcomer the shape of a good solution well before they could produce one themselves. And it compresses research into short, pointed bursts, which is what lets one person cross from code to production the same week.

Depth: A Feature a Week

Rohit spent the summer inside the core of our AI review product, and 32 of the merged pull requests are his. Up until this summer, when our AI finished reviewing a submittal, it handed the reviewer a list of findings and stopped there. Rohit built the step that weighs how severe those findings are and turns them into an overall recommendation: approve, approve as noted, or revise-and-resubmit.

He followed with a pipeline that cross-checks the products in a submittal against the project’s equipment schedule, which is a classic place for a mismatch to slip through.

He also built collaborative annotations so several reviewers can mark up the same PDF and see who wrote what, and tightened our runtime configuration and production log search.

His own summary is that with the routine code handled, his time went to whether the thing actually works, what should happen when it goes wrong, and whether the output is useful to someone in construction. “It feels a lot more like the work a developer would do than a typical internship.”

Breadth: From Market Research to Code

Kenley’s summer proves that the AI landscape allows for breadth as well as depth. This was her very first internship, and she started by running live submittal reviews for customers on projects with real deadlines. She noted every bug she found and every opportunity for UX improvement. Within weeks, she was pushing her own pull requests.

In addition to ideating and executing UX improvements, she audited our entire runtime configuration, cataloguing every variable in the platform and finding the minimum it needed to boot (which was the foundation Rohit later used for his cleanup). For other tickets, she reviewed what Cursor produced, tested it against the acceptance criteria she had written, and sent back what failed to hold up. That skill, judging output you could not have produced from scratch, is the new entry-level skill, and Kenley used it to great success.

Her biggest project this summer was revamping all of our investor-facing materials and building a construction news agent. She rewrote our entire data room, edited both the investor deck’s text and design, and headed the market research behind it. She also built a Claude Cowork agent that systematically scans reputable sources for the latest construction news and curates the findings to keep our team updated.

By supplementing her research skills with AI tools, Kenley was able to own truly consequential work across multiple fields despite being an entry-level hire.

The Part That Stays Human

All of it rests on judgment, which is the thing we actually screen for when hiring. AI can confidently produce wrong answers, and the person who forwards them can spread the errors. The guardrails exist so a hallucination dies in review, but someone still has to catch it, ask whether the answer is even relevant, and own the call.

Which brings me back to how we teach entry-level work. We used to teach what we had learned ourselves, coding, building decks, doing the work. Now we teach spotting inconsistencies, explaining what good looks like, defining what done means, and recognizing the telltale signs of a hallucination.

The old internship deferred judgment for years and hoped it would arrive on schedule. Everything shifted from doing the work to exercising good judgment.

We had to change our habits deliberately, including plenty of what used to be called best practice, to make the most of the new reality. We are better for it, and I’d like to think our interns got a much richer, more meaningful learning experience out of it too.

I hope sharing how work has evolved at our company is useful to other companies out there. I am always happy to share insights and unlocks if anyone wants to know more.

But above all: thank you, Kenley and Rohit, for an awesome summer of work!