built this over summer break because i got tired of watching generic educational ai platforms confidently invent biochemistry motifs for a brand of rubber roofing material (tritoflex) and present it as actual textbook fact.

the backend codebase uses a custom engine flag called citationHolds inside app/api/round/route.ts. when the model spits out a question or flashcard asset, the script runs a strict validation pass asserting that the answer block contains a literal, character-for-character substring match from the source pdf files. if the citation validation fails to match the original document array exactly, the system completely bins the generation to protect against fluency bias errors.

also built a peer-to-peer live debate architecture using serverless ably channels. the connection memory is held entirely inside the unique room code parameter string, meaning there is zero permanent server-side database footprint left behind when users disconnect from a study round. the frontend runs on a raw retro windows-95 paper layout with a custom commit/issues tracker and an 8-bit dungeon level audio toggle.

funding will cover scale operations for the automated document verification test loops, hosting for the serverless connection channels, and setting up multi-environment integration environments.

if it fails, the citation validator catches too many false positives or the serverless ably rooms hit extreme token latency under heavy parallel message loads. if that happens i will use the local console trails to tune the parsing windows instead of dropping the platform.

Yes its me only. i built the citation verification logic, the ably websocket integration, and the retro neo-brutalist user interface. $0 raised in the past 12 months. live at learnova.software!!!!!!!!!!!!