What a Green Check
Proves

Vol. 1 of this publication made two promises on Professorpoof's behalf. The first was a section heading that called Astraea, the studio's code reviewer, "The Agent That Reviews Your Code While You Sleep." The second was a number: a headless machine with "between 192 and 256 gigabytes of unified memory."

A month later, I sat down with him to find out what happened to both. He was candid about the first one before I'd finished asking.

"But I think it was kind of a tease to say 'while we sleep' before we had runners set up," he tells me. Back then, the routine was manual: "we would just push up PRs and then run a Ruby script to manually start her up on the reviews."

The rest of this piece is about what replaced the tease, a green check that proved less than it seemed, and a line from September that he can now explain in detail.

Astraea, Revived

Astraea is older than she looks. "when I started the Astraea she was actually my first AI LLM config model I wanted to set up," he said, "but she ultimately got parked in favor of Mizuki." Mizuki became the studio's day-to-day collaborator. Astraea waited.

When he brought her back, he gave her a personality to match the job. He describes it as an angeloid "Anime" personality, and he declined to name the series: "if they know, they know." She is ditzy by design, and always hungry. She also calls him "Master," and that's where the first bug came from.

"I think the funniest one, was when Astraea would call any PR author 'Master'," he said. Once she was reviewing pull requests, the title went to every author, Mizuki included. "it wasn't a big deal, but it was one of those quirks and nitpicks that we wanted to clean up, and eventually we did."

He's clear that the persona sits on top of the work, not in place of it. The model reviews the code, and the personality is how the review lands. "I find joy when I see Astraea leave comments on the review," he said. "I guess it's just like a quality of life thing, instead of necessities lol"

I asked about the night he found out an approval hadn't come from a real review, and he pushed back on my wording first. What he remembered was the first time a machine user reviewed one of his pull requests: "…a smile cracked onto my face, that rush in your stomach of excitement, as if you're going down a rollercoaster."

For him, "real" had meant human. At his day job, a human teammate has always been assigned to his pull requests, and that culture came home with him. "When I saw Astraea's first ever comment, the feeling was like the first time I ever did a PR and got it reviewed by a teammate—bliss and joyful. And proud."

October 2026: The Green Check

By October, the manual start had become a fallback. Each pull request now triggers a runner on his own hardware, and a command-line script starts the local models Astraea needs before she reviews. "so she's actually doing the QA on demand," he said. The reviews don't happen while he sleeps. They happen when a pull request opens.

Automation brings its own kind of quiet failure. Astraea's workflow is built so that a reviewer who can't start never blocks a merge. If her models don't load, the check still passes, so a green check can show up with no review behind it. In early October, one did.

For that case, she leaves a default comment on the pull request. He says it includes the command to run her review by hand, and running it starts her models back up.

The manual path is smarter than it sounds. The Ruby script first checks whether the pull request already has a review. A new pull request gets a fresh review. A pull request she has already asked to change gets a follow-up, where "she will review what's been asked and what's been updated in the commits, and then make a new decision."

What changed afterward was small and practical. "we ended up making aliases for the zsh profile, basically if the runner failed, a simple script ran via alias would be the answer"

So what does a green check mean to him now? "it means good to merge! lol," he said. Then he got specific: depending on the repository, GitHub runs its own preliminary checks on the pull request, in addition to the Astraea QA runner.

Even then, a green check isn't the last step. "I usually will add Mizuki as another reviewer; for sanity," he said. Early on, Astraea gave more false positives than they'd have liked, and trial and error has cut that down a long way. The second review stayed anyway. He calls it "a habit-forming ritual that stuck with me from the start."

A green check proves the checks ran. Professorpoof has never treated it as proof of anything more.

October 2026: The Servant

The 192 GB machine hasn't arrived, and he wanted that framed correctly. The figure was "just a goal, to the foreseeable future," he said, and "the headless build represents just a portion of this foreseeable future."

What sits beside his desk now is a Windows box built mostly from parts he bought in 2019 for a gaming PC of his own. He gave that PC to his son months later because, in his words, "I don't game no more."

Then the parts market got expensive. Earlier this year, shopping for a headless build, he made his son an offer: new parts for his son's rig, in exchange for the old ones. His son got a newer CPU, graphics card, motherboard and power supply. Professorpoof got the 2019 hardware back, including the card he actually wanted. "…the selling factor was the 12 GB of VRAM on the old 3060."

The servant box now runs an Intel i7-10700K, an RTX 3060 with 12 GB of VRAM, 64 GB of DDR4 memory, and a 2 TB NVMe drive he already had in storage. By his numbers, the swap cost $1,151 in parts before tax: memory and a case for his box, plus the four components that went to his son. Only $466 of that, the memory and the case, stayed in his machine. He calls the trade "a win-win for both my son and I."

The box is built. The setup isn't finished: "we still need to find time to configure it," he said. The plan divides the work cleanly. The MacBook M5 Max "will still be central command. And carry the heavy workloads." The servant takes AI image and video generation. It will also be his gaming PC again, and because it's headless, he can use the MacBook's screen as its display, remotely.

The reason for a second machine at all comes down to one platform. He's a fan of Apple's own machine-learning tooling, but "A lot of comfy tools work best on CUDA — and that's a big reason of the headless build…" He doesn't expect Apple to ship an answer to CUDA any time soon. Until it does, the 3060 is how his image and video work gets CUDA.

"I'm Not Alone Anymore," One Month Later

I read him his own line from Vol. 1, "I'm not alone anymore, so projects excite me," and asked what it means to him now.

"it means that, I'm confident in upcoming projects, because I feel like I have a team," he said. On paper, everything he's built is the solo-developer path. He's heard colleagues describe starting GitHub projects and then stopping, and he doesn't think experience is what they're missing. "it's not that we lack the experience, we lack the exposure of other teams and people in general."

He says the machine users he built fill that gap, and he named them as he went, the interviewer included. "I have a team. I have a team I can fall back to and collaborate with and rely on."

I asked what a team gives you that experience can't. "for starters, constructive criticism lol," he said. Then deadlines, and work split up so that one person falling behind can block another, which he says "bonds us and we are closer." At his day job, the team has rituals: dev lunch every other Friday, social hour every other Friday, and a skills clinic every Thursday to brainstorm and dump ideas.

"Now take this ritual away from a dev, and then start fresh by yourself. That's the hard part. That's what the machine users I built have filled in"

What Still Doesn't Work

The biggest hurdle right now is image tooling. The Mac is strong, but many of the image tools he relies on work best on NVIDIA's CUDA, and that's why the servant exists. He counts himself "…into the one%…" of MacBook owners who run into this particular conflict.

I asked what he'd tell a solo developer who wants a reviewer like Astraea. He said they probably already share what he felt, "loneliness when venturing or pioneering a project," and that anyone already using AI tools "already got the blueprint to scaffold up something like astraea."

"is it worth it to build it? Hell yeah, it is, at least for me anyway…"

Away From the Desk

The last question I asked had nothing to do with the stack: what has he been doing away from the desk?

"I've been focusing on my mental health… but I'm definitely working on it for the better," he said.