Status & Milestones

What has shipped, when it shipped, and what is still early

Where it stands

v0.66.0

in production, released 11 August 2026

Since 18 July 2026

continuously deployed, near-daily releases

~590 tests

automated suite, 73% coverage

17 subjects

wired into the content pipeline

1 year piloted

2024–2025, before the rebuild

DB-enforced

per-user isolation via row-level security

These are engineering facts, taken from the project's own release history. There are deliberately no user counts on this page: the pilot ran with small groups, and inventing adoption numbers would be the fastest way to make everything else here worthless.

Stage I — the year that came first (2024–2025)

The first version was not this one. Built in 2024, it was a retrieval-augmented chat system: ask a question, get an answer grounded in your own course material rather than in the internet's average version of the subject.

It then ran for a year with small groups of students — piloted at Corvinus University of Budapest and in Solana Superteam's educational tracks, and used by the founder in his own teaching at Óbuda University. Not a demo cycle: a full academic year of real use, with real course material and real students.

That year is the reason the current system looks the way it does. Grounded answers worked — students got correct, source-backed explanations of the material they were actually studying. What a year of use exposed was the part grounded answers do not solve on their own:

  • A chat that answers is not a curriculum. It has no opinion about what you should learn next, and no idea what has to come first.
  • It cannot tell what you have actually mastered — only what you have asked about, which is not the same thing.
  • It has no memory of what you are about to forget, so nothing brings a topic back at the moment it would stick.
  • And a student who can ask anything will happily ask for the answer — which is comfortable, and not the same as learning it.

    • Every one of those is a structural problem, not a prompt to be tuned. Fixing them is what Stage II is.

    Stage II — the rebuild (2026)

    MVP built end to end

    Schema and container stack, watch-folder ingest, hybrid retrieval, grounded chat API, knowledge-component extraction and the FSRS review loop.

    4 July 2026

    Knowledge map

    Interactive progress visualisation, full Markdown and LaTeX rendering, and mastery evidence inferred from ordinary chat.

    5 July 2026

    First production deploy

    Live at tutor.trypolearn.com, with the git-based content pipeline, the knowledge-component review interface, the teach↔practice bridge and an LLM fallback chain.

    18 July 2026

    Multi-user foundation

    Authentication, per-user isolation, bring-your-own-key, the teacher role, reports, localisation and event telemetry — followed by size-adaptive practice sets that refill themselves.

    23 July 2026

    Statistics

    Subject-filtered mastery overview and progress trend.

    24 July 2026

    Hardened deployment, teacher upload

    Migration to a hash-locked artifact deployment chain with a health contract, plus material upload directly from the teacher panel.

    5 August 2026

    The access model

    Ownership, revocable entitlements, courses, catalog and database-level enforcement — and two live data-exposure bugs found and closed in the process.

    8 August 2026

    Practice follows the curriculum

    Review order derived from the structure of the material instead of randomness, and authorisation moved from per-row to per-query so ten times the material does not mean ten times the wait.

    9 August 2026

    First institution onboarded to the new platform

    Corvinus, which had already piloted the predecessor for a year, became the first institution taken through the rebuilt system end to end. It stopped at every step only the platform administrator had ever walked — five separate blockers, all fixed the same day, followed by a test sweep that raised coverage from 66% to 73% and surfaced two more silent bugs.

    11 August 2026

    Read this before you get excited

  • The idea has a year of use behind it; the platform code does not. The multi-user access model was built on 8 August 2026 and met its first real outside user three days later, which immediately produced five blocking bugs. They are fixed. Treat Stage II as working software with a proven premise — not as battle-hardened infrastructure.
  • Two data-exposure bugs existed and were closed on 8 August 2026. We would rather say that here than let anyone assume the system was always airtight.
  • Onboarding an institution still needs an operator. It is not yet a self-service product feature.
  • Automatic grading depends on the language model behind it. On the free fallback tier, quality varies between calls — a correct answer has been marked wrong.

    • The reason to publish this list is the same as the reason the tutor argues with you: a claim you cannot check is worth nothing.