LearnCoachAssist

Spaced repetition, with AI deck generation that can't quietly bill you for nothing.

Type
Product
Industry
EdTech / education
Platform
Web application
Role
Design, build, deploy
LearnCoachAssist spaced repetition study screen

What made this hard.

A study platform built around SM-2 spaced repetition, extended with an LLM pipeline that generates whole decks from a topic, plus quizzes, duels, public decks and Stripe-backed credit bundles.

Spaced repetition is well-understood; the differentiator was generating decks on demand. That is where it got interesting, because model calls are slow, occasionally truncated, and always cost money.

Doing generation inside the request meant timeouts on large decks. Charging credits up front meant users paid for failures. And a second study code path for generated decks would have doubled the surface area of the part of the product that has to be correct.

What we built.

  • Moved deck generation out of the request path into a queued job: status tracked on the deck row, credit reserved at submit and idempotently refunded on failure, an ETA returned immediately, and an email when it completes.
  • Added a per-user dedup hash so an identical request returns the existing deck instead of paying for it twice.
  • Broke one huge model call into fixed-size batches, each steered away from questions already generated, merged and de-duplicated on a normalized question key with per-chunk retry budgets, bounding the blast radius rather than widening the timeout.
  • Wrote salvage parsing for truncated JSON arrays, so a partial model response still yields usable cards instead of nothing.
  • Decided by ADR to reuse rather than duplicate: user-generated decks render through the same study view as catalog topics via a synthetic payload, and SM-2 progress tracking was extended to accept them.
  • Root-caused a flapping test suite to a config-cache leak (a cached config froze production drivers and silently overrode the test environment) and fixed it at the PHPUnit bootstrap so the guarantee holds from CI, cron, an IDE or a shell alike.
  • Added Labs, guided Skill Paths that run beside the flashcards: 16 paths, 203 units and more than 1,000 lessons. Math answers are computed by tested code; the other paths are written with AI and checked by automated tests. The free tier gets 20 new Labs lessons a day and paid plans are unlimited.

What it runs on.

Chosen to be maintainable for years, not to win an argument on the internet.

  • PHP 8.2+
  • Laravel 12
  • Tailwind CSS
  • Vite 7
  • MySQL
  • Stripe
  • OpenRouter
  • Database queue
  • systemd

Where it landed.

The product ended up broader than the flashcards it started as: flashcard, type-answer, learn and match study modes, quizzes, duels, public decks, profiles with avatars and bios, referrals, and credit bundles with a Stripe trial, all on one study engine rather than several.

Labs is the newest part: a free-to-start set of guided paths, 16 paths and 203 units with more than 1,000 lessons, 20 new lessons a day on the free tier and unlimited lessons on paid plans. It is new, so we do not quote usage numbers. You can try it at learncoachassist.com/labs.

Because generation is queued and credits are reserved rather than spent, a failed model call is a refund and a retry, not a support ticket.

Need something in this territory? This is AI agent & LLM development work. Start from all services or see pricing.

Your project

Tell us what you're building. We reply within two business days.

Hour Blocks from $180reply within two business days Start a project