The criteria were in the teacher’s head. That was the problem.

Students submitted work — text, photo, video, slide decks — into three Telegram bots running on n8n, Supabase and Gemini. Domain logic was scattered across n8n nodes, the free tier throttled the project, the grade log lived in Google Sheets, and neither student nor teacher had a single interface. The teacher had no time to review every submission by hand and explain what each one needed to fix.

One app replaced all three bots. A student submits work in any format and is released immediately while review runs in the background; if the work is rejected, they get a concrete numbered list of what to fix, each item referencing the criterion it fails. The teacher sees a feed of incoming work with the review result already attached, and gives the final grade. The bots’ domain logic was ported 1:1 into an Express API with raw SQL over PostgreSQL on the client’s own VPS — no third-party limits, no credentials in the client bundle. Three problems needed real engineering: an atomic work-claim so a fast submitter can’t burn their attempts; a DB-persisted retry queue that survives restarts under load; and a move to a Tashkent VPS with a reverse proxy after the review service refused certain server locations. A pseudonymous ranking with per-viewer animal masks keeps the score gap visible while hiding who’s behind.