Stories from teachers.

I am not a person who scans things. I'm a first-grade teacher. My day starts before the sun and ends with my forehead pressed against the steering wheel. The scanner is at the office, and the office is two hallways and a printer jam away from my classroom.

When I tried HomeworkGrader the first time, I was sitting on the rug at 3:14 p.m. with twenty-four math papers in my lap. I opened the site on my phone, took a picture of each paper one after the other, and tapped Grade. By the time I had packed up, I had the results sorted by problem.

Six months in, I have not opened the school's scanner once. The flexibility is the whole product for me. The AI is amazing — but the fact that it works on whatever I already have in my hand is what made it stick.

By the time you are on paper twenty-three of thirty, your eyes are doing the work. You stop reading; you start matching shapes against an answer key. I know I do it. I have checked my own grading against a colleague's enough times to know I do it, especially after a long day.

In early March I ran a stack of thirty fraction worksheets through HomeworkGrader and it flagged a paper I had already marked. Problem six was a fraction I had checked off as correct: 2/6. The right answer was 1/3. I had seen "2/6" written in the same neat handwriting I had been seeing all afternoon, and my brain filled in the rest. The grader caught it because it does not get tired.

I do not think of this as the AI being smarter than me. I think of it as a second pass that does not have to drink coffee. I would rather a student get the right grade because two graders looked at the paper — even if one of them is software.

I have been an English teacher for twenty-two years. The job has had exactly one constant the whole time: grading happens after dinner, and it ends when it ends. For most of my career, "when it ends" was 11:30 p.m. or later on the nights before grades were due. Five-paragraph essays are not a problem you can run over.

I started using HomeworkGrader in late January for the Monday opener — a short, twenty-minute reading quiz I use to check whether the weekend's reading actually got done. It used to take me forty-five minutes to grade twenty-eight of those during my prep period. Now I run them through the grader during the bell change between first and second. The whole thing takes under four minutes.

The bigger essays still take time, but the AI now drafts the verdicts and feedback, and I read only the contested ones — usually about five out of twenty-eight. My nights have started ending at a normal hour. My partner pointed out, two weeks in, that I had watched a movie with him on a Tuesday for the first time since 2009.

When we first piloted HomeworkGrader on the writing side of our class, we expected math-level accuracy on math but a noticeable drop on essays. Writing is judgment work, and judgment is hard to automate. We ran a three-week comparison study: every short-response answer on every quiz got graded twice — once by the teacher, once by the AI using a rubric we wrote and pasted in.

Out of 1,184 graded responses, the AI agreed with the human verdict on 1,180 of them. Two of the four disagreements were cases where the teacher had been generous; the other two were genuine edge cases where the AI was too strict on a partial-credit answer. We adjusted the rubric to allow more flexibility on supporting-detail wording, ran another batch, and the disagreement rate dropped to under one in five hundred.

The lesson for our department was not "the AI is right." It was "the rubric runs the show." The same rubric, applied consistently across thirty papers in three minutes, produced grades that were more uniform than what four different teachers produced the week before. We have since started using the grader's output as a calibration check on our own grading meetings.

Three days before winter break, the seventh-grade math department at our middle school administered the same end-of-semester quiz across four sections. By the time the bell rang Friday afternoon, the four of us had a stack of 142 papers between us — and most of us had report-card deadlines on Monday morning.

We pooled them on one table in the department office and ran the entire stack through HomeworkGrader in a single batch. Each teacher got back per-student verdicts grouped by problem, sorted by score. We split the review by section and worked in parallel for forty-five minutes. By 4:15 p.m., we had reconciled the grades, recorded them, and gone home.

What used to be a guaranteed weekend of grading became a ninety-minute coordination problem. The win was not really speed — it was that the work fit inside the workday. Three of us went out to dinner that Friday. The fourth picked up her kids from school for the first time in a month.

Saturday mornings were grading mornings. Twenty-seven math papers, the kitchen table, two cups of coffee — a ninety-minute ritual that ate the best part of the weekend. Anjali, a fourth-grade teacher at a public elementary school in Newark, had been doing it for nine years. "I was starting to resent the work itself," she said. "Not the kids, not the lessons. Just the stack."

She tried HomeworkGrader on a Friday afternoon in late September, expecting to spot-check what the AI got wrong. She sorted the results by score, ran her eyes down the wrong ones, agreed with every verdict, and was done in twelve minutes. "I kept waiting to find the catch," she said. The catch never came. She walked out of the building with the grading book closed.

Three months later, the routine has stuck. She still reviews every paper — she is not handing the work over, she is accelerating it. Saturday mornings are now a long walk with her dog. "It is such a small change," she said, "but it is the difference between a job that ends and a job that follows you home."