TrueMark

Pages

Teachers grade in pages. Our editor scrolled. So teachers sent students to Google Docs, and schools wouldn't pay for a writing tool their teachers routed around.

I put Word-style pages inside the editor: breaks as you type, editable headers, page numbers, footnotes that stay on their page, print that matches the screen. Started from one open-source extension, rebuilt most of it.

The Google Docs objection stopped coming up in sales calls. 130,000+ essays. TipTap shipped the same feature 6 months later.

The tutor

The tutor is meant to ask, not answer. One teacher said it was writing for her students. One opinion can't change a prompt 45,000 students use, so I measured it.

Pulled 186 of her tutor conversations and 200 matched ones from other classes out of Langfuse (where every AI conversation is logged) and had a model grade each turn: did the tutor write something the student could paste in? Hers: 13.0%. Others: 5.5%. She was right.

Fixed the prompt in a day. The same replay now runs before every AI change, and the PM runs it herself. A student's review, from the TrueMark homepage: "Your copilot sucks it didn't write my essay."

Exam day

In June, finals for 14 sections hit submit in one period, about 35x what the product could take. Every submit generated its report on the API pod, and compose_diffs spawned uncapped threads that starved the event loop. Health checks timed out, Kubernetes killed pods faster than they came back, 8 of 12 died. 158 students got errors mid-exam. Districts don't renew after a day like that.

Split submit from grading. Submit now writes the submission and enqueues a Cloud Tasks job; a Cloud Run worker pool generates the report, with an inline fallback if the enqueue fails so the new path can never be worse than the old one. Deleted compose_diffs for a page-length index that is binary-searched at read time. Rolled out behind a flag, 0 to 100% over 8 days, alerts watched at every step. Load-tested by replaying a whole district: 5,000 students, 4,175 of 4,175 submits, all 292,617 Firestore writes verified.

Two months later, a district benchmark 30x the size of the June class: 4,338 essays in a day, 2,844 students, a peak of 44 submits a minute (a 2,600-an-hour pace, 1.5x the rate that killed it in June), 0 failures a student saw. The June district renewed. Back-to-school went through, in the team's own words, "without depleting customer goodwill."