Do You Have the Vibe?

Near the end of the semester I asked our IT department to keep the exam-scoring software running through finals. They declined, having found it was no longer supported. The first move was to ask. After that was refused, I still needed digital answer data to give students individualized exam reports, so I decided to build it.

Student responses could not be uploaded to an outside service, which ruled out the convenient answers and GenAI alike. And reliable was not the whole requirement. I needed something I could audit myself, so that nothing went out as a grade until I had checked it. All I actually had to obtain was the digital answer data. That decided the shape of the tool before a single feature did.

An optical mark reader for a custom 100-question sheet, running entirely in the browser with Python compiled to run there rather than on anyone’s server. It registers each page from its four corners, adapts its threshold to the scan, and catches multiple marks. Before any grade goes out it flags blanks, double marks, and low-confidence reads, so a person reviews the doubtful ones first. A guided roster step at the front means every sheet arrives already carrying a name, and captured page images stay in the session rather than being stored.

I did not need a perfect replacement. I needed a local, testable, good enough one.

The reader solved the scoring. It did not solve the reporting. I was still exporting a CSV, opening an Excel template I had built, updating several parts of it by hand, checking the result, and only then printing something a student could learn from. It worked. It also meant the job was spread across two tools and a stack of manual steps every single time I gave an exam.

The report moved inside. The same app that reads the marks now assembles the individualized report, so the path from a scanned answer sheet to the page a student holds never leaves one file. Each report carries the student’s answer against the key for all one hundred questions, the chapter, cognitive level and learning objective behind each one, and their standing by all three of those against the class. Twenty-four students, ninety-six pages, one click, padded so a two-sided run starts each student on a fresh sheet.

An emergency replacement, left alone long enough, turns into the thing you actually wanted.

Getting here was not tidy. Versions broke things that had been working. Ideas that looked trivial turned out not to be. The work was less about writing the feature than about noticing, each time, which part of my own process I had never actually written down.

Every term ends with three required files from every instructor, for every section they taught. Collecting them was never the hard part. Faculty submit through a form and everything lands in one place. Knowing who is still outstanding was the hard part. It meant opening the shared folder, filtering by a faculty name, finding a file, then clicking it to confirm which section it actually was, and starting again with the next name. I built an Excel workbook to cut the clicking down. It helped, and it had to be reformatted at the start of every term, so the workaround became its own small chore.

The form does not record which section a file belongs to. The only clue is the file name, and file names are written by people, so the same submission arrives as Alvarez_BIOL_1406_86002_CAN.pdf one term and Alvarez-BIOL-1406-86002-CAN.pdf the next. There is a second trap underneath that one. Section numbers are not unique across the college, so a biology section and a humanities section can carry the same five digits. Matching on the number alone is safe only when the submitter is already on my roster.

The app takes the roster on one side and the form’s recording sheet on the other, then works through the file names. It reads the five digit section number, corroborates it against the course number in the same name, keeps the most recent version when a section is sent twice, and ignores everything before a date I set so last term’s rows never count toward this one. What comes out is the list I used to assemble by hand. Every instructor, every section still open, and which of the three files each one is missing, ready to copy straight into a reminder email.

The useful move was making it decide less, and making it say out loud where it stopped.

Deciding when to offer nothing at all. When two of an instructor’s sections are equally close to the number in the file name, the app suggests neither. Two sections one digit apart are the easiest thing to mix up and the worst thing to get wrong, so a blank is safer than a confident answer. Everything it cannot place arrives with its reason written beside it, which is the part that actually saves the time.

Teaching Squares asks four instructors from four disciplines to visit each other’s classes across a semester. Nobody is evaluated, and the value is in watching someone teach a subject you do not teach. The obstacle is that all four of them are teaching. A colleague’s most interesting class meets at the hour you are standing in front of your own, and the only way to find out is to lay the schedules side by side. For a cohort of eight that is tedious. For a cohort of sixteen it stops being something a person should be doing at all.

Observation is a two-way relationship, so a Square of four holds twelve separate questions. Whether A can visit B is a different question from whether B can visit A. Sixteen people divide into four Squares 2,627,625 ways, and a grouping that looks balanced on paper can still contain one person nobody in the Square is free to visit. Of those 2,627,625 arrangements, exactly 16 leave twenty-three of the twenty-four pairs workable in both directions. Nothing reaches twenty-four. A random arrangement averages under sixteen workable pairs.

It imports the college’s own schedule export, so nobody retypes a week. The import applies the rules that make a schedule readable. A cross-listed section appears once rather than twice. A dual credit high school section counts as teaching time that cannot be visited. A meeting with an incomplete time is reported rather than guessed at. Then it tests every ordered pair. A class is visitable for a given observer only when that observer is free for the whole class plus a transition buffer, so a fifteen minute walk across campus does not turn into a late arrival. Each Square that comes out of the search gets a weekly calendar image, color coded by person, with rooms, times and the possible observation weekdays printed underneath.

A search that reports what it could not solve is worth more than one that returns a tidy answer.

Deciding what the app is allowed to settle on its own. The facilitator unchecks labs, clinicals and studios by hand, because judging what is worth observing belongs to a person, and an unchecked class still counts as busy time for the instructor who teaches it. In the run below, one direction out of forty-eight has no complete class visit. Nadine teaches calculus at the exact hour Dean teaches his only visitable class, because his studio section was withheld from observation. No arrangement of this cohort fixes it, and moving either of them breaks something else. The app names the pair and leaves the choice open. Every option it produces is a weekly possibility rather than an appointment, since exams and presentations stay with the instructor and the visit itself is confirmed by email.