An AI teaching workspace that carries a lesson from standards through Google Classroom to returned grades.

A competitive scan of MagicSchool, Brisk, Khanmigo, Curipod and Diffit showed the category racing on breadth, time-saved claims and trust badges — a race a new product can’t win. The positioning work aimed to change the axis of competition instead: a persistent planning workspace, quality that can be shown, and depth on standards rather than breadth.

Walking the product through real teaching tasks before pilots produced a batch of changes: a preview step was cut so generated lessons save directly, adapting a lesson moved onto the Home screen where a teacher starts the day, coaching-style review was split from scored audit, and a setting that could silently delete lessons was removed.

Conversations with teachers informed the product throughout. One documented example: lessons carry a movement attribute because a teacher described “my after-lunch group — they need to be up and moving.”

Objective questions score themselves. For written work the model drafts a score and rationale and stops there — it can decline with “needs teacher judgment,” and never sees student identity. Grading is a judgment a teacher answers for; automating it would move accountability to something that can’t hold it.

VeraTeach could have drifted into becoming an LMS. Instead it prepares, publishes and grades, while Google Classroom stays the system of record schools already run on — which is also why the Workspace Marketplace listing was pursued for district admin allowlisting rather than discovery.

A broader search would surface more standards. Retrieval stays bounded and the model abstains rather than padding the list, because in a standards-aligned product a confidently wrong suggestion is worse than a narrower right one.

Licensing a standards API would have been faster. Owning the ingestion means alignment is verifiable, adding a state costs days rather than a negotiation, and no vendor sits between a teacher and whether their lesson is correct.

Plans launched as hypotheses, not convictions. Downgrading deletes and locks nothing, and pricing copy was corrected whenever it promised something enforcement didn’t. When an early free plan turned out to allow a complete, forever-free workflow, limits were reset so free stays useful without replacing paid. The launch offer was a one-time discount, not a price, so it couldn’t quietly become permanent.

Letting visitors generate a lesson before signing up was built to lower the barrier to trying the product. The data showed it wasn’t converting and was the only unmetered AI spend, so it was retired in favor of signing in first and starting immediately after.

A hard filter on state, subject-eligible frameworks and grade; vector similarity over the embedded standards; then an LLM rerank that can only choose from the retrieved candidates, with a confidence gate that abstains when nothing is a strong match. Results are cached against the corpus, model and prompt versions.

I built and maintain the corpus itself — 185 frameworks across 17 states, 110,000+ records — ingested from each state’s own documents, every one with its own format and parsing failures.

Generation, review, adaptation, materials and grading suggestions all run through a single layer that enforces a typed schema and a timeout, and records model, latency, tokens and cost for every call — which is where routing and cost decisions get their data.

Rubric evaluation

Lessons are scored on eight explicit criteria, from objective clarity to teacher usability, and any evidence the evaluator cites must appear verbatim in the lesson, or it is dropped.

Blind model comparison

Model and prompt variants are compared blind, with output order shuffled, on quality, latency and cost. On generated tests, a three-way comparison found equal quality at ~33% lower cost and set the routing for summative assessments; low-stakes materials run on a model roughly 10× cheaper per token.

A measured model change

Moving rubric evaluation to a cheaper reasoning model cut it from 74s to 19s at ~2.1× lower cost, with a mean score difference of 0.05 on a 0–4 scale across a fixed lesson set.

A benchmark that rewarded the wrong thing

One evaluation benchmark turned out to reward confident invention over asking a clarifying question. I changed the benchmark before trusting its results.

Regression, retrieval and corpus checks

Live checks run against real models when a prompt or model changes, across a baseline matrix of K–12 subjects and grades. Retrieval recall is benchmarked, and every standards framework passes coverage checks against its source before teachers see it.

When traffic wasn’t converting, I investigated before changing the page: most catalog “views” turned out to be bots, and real visitors were largely mobile ad clicks looking for things the catalog didn’t cover. The first fixes were to the measurement, not the product.

A unit-first homepage was reverted the same day, once it was clear it met lesson and quiz searches with a page about something else. A landing change that stopped ad signups was rolled back within two days.

Every creation path once failed for about 14 hours behind a friendly error message that no monitoring saw. The lesson: a user-visible failure that degrades gracefully is invisible to every system you run — so alerting now watches what users experience, not just crashes.