The platform is 19 small, focused repositories under @EvilFlowersCatalog — a catalog at the centre, a fleet of workers around it, one protocol tying them together, and the student teams who built the whole thing. Jump to a layer:
Start here. The four pieces you would actually run — the catalog that holds everything, the portal people log into, the reader they open books in, and the importer that fills it.
OPDS publication catalog server. Stores publications, exposes them over OPDS 1.2 (Atom) and OPDS 2.0 (JSON), and hosts many libraries from a single install. Django backend, Pydantic schemas, optional Readium LCP lending.
React + Vite frontend. Catalog browsing, search, user shelves, university SSO. Single codebase, themed per deployment.
Embeddable PDF viewer built on pdf.js. In-document search, citation export (BibTeX/BibLaTeX/RIS), annotation layer, page-range sharing, theming. Ships as an npm package.
Imports publications from Kramerius and WebDAV sources, then uses an LLM (OpenAI or Ollama) to fill in missing metadata — titles, authors, abstracts. Outputs JSON the catalog can ingest.
The heavy lifting. Independent services on a job queue that read every publication — pulling text, captioning figures, reconstructing equations, transcribing video, encrypting loans — so the catalog can stay simple.
Celery worker. Extracts text from PDF publications and emits it back to the catalog for indexing and downstream analysis.
Semantic search over the extracted text corpus. Sits behind the catalog search endpoint and consumes output from the text service.
Celery worker wrapping Readium lcpencrypt. AES-256-CBC encryption for EPUB 2/3, PDF, Readium Web Pubs and W3C Audiobooks. Filesystem storage only.
Extracts figures from publications and runs captioning (PaliGemma) and classification (ViT / ResNet50). Outputs captions and tags for accessibility and image search.
Detects equations in PDF publications and reconstructs them as LaTeX using the Nougat model.
Transcribes audio tracks from video publications using Whisper large-v3. Output is searchable text aligned to timestamps.
Pipeline orchestrator. Routes a newly ingested publication to the text, image, equation and video services and aggregates their results.
Dockerised Celery worker wrapping ocrmypdf. Adds a searchable text layer to scanned PDFs ingested by the catalog.
The contract. One OpenAPI definition of how the pieces talk, plus the generated client so you don’t hand-write requests.
OpenAPI 3.1 definition of the EvilFlowers protocol — the wire contract between catalog, portal, importer and third-party clients.
Python client and Pydantic v2 models generated from the protocol spec.
The relay. Each FIIT STU student team leaves a repo behind — features and forks defended as thesis work, then handed to the next crew.
FIIT STU Team 23, 2025/26 academic year. Current student squad.
In-portal chat assistant. Searches the catalog, summarises publications and answers questions about library holdings.
Fork of the text service used by the 2024/25 team. Frozen at the version their thesis was defended against.
FIIT STU Team 5, 2024/25 academic year.
FIIT STU Team 16, 2021/22 academic year. Where ELVIRA started.
Deployment manifests, runtime forks, benchmarks and the external standards ELVIRA implements don’t get cards here — they live on the ecosystem map and in the org. That’s where the wiring is.
