Reviewer2.0 reviews an academic paper (PDF or pre-parsed JSON) with a panel of independent AI reviewer agents — one per rubric, choosable per review — each reading the sections it needs (and, for citations, checking them against the real literature) before scoring it 1–10 with written justification and bullet notes. A supervisor agent then synthesizes the panel's findings into an overall recommendation. Reviews can be exported as Markdown or JSON.
Originality, Soundness, Impact, Presentation, Positioning w.r.t Related Work, Reference & Citation Quality, Reproducibility & Artifacts, and Ethical Considerations & Broader Impact. Pick any subset per review.
backend/— a FastAPI service. Each rubric is a LangGraph tool-using agent (create_react_agent) that decides for itself which paper sections to inspect; the reference-checker agent additionally verifies citations against the real literature via a user-selected lookup provider — Tavily web search (default) or Semantic Scholar. Selected agents run in parallel, then a supervisor LLM step synthesizes their findings. Results stream back over Server-Sent Events (GET /api/criteria,POST /api/review) so the frontend can show live per-agent progress instead of one long wait. PDF uploads are converted to structured text (and a parsed bibliography) via GROBID; pre-parsed JSON uploads ({"markdown": "..."}) skip that step.frontend/— a static HTML/CSS/vanilla-JS app served by nginx. It lets you pick which rubrics to run, streams the live review panel's progress, and renders the results (with score gauges per rubric) — Markdown/JSON export is built client-side.grobid— the GROBID PDF-to-TEI-XML service, run as its own container.
cp .env.example .env # fill in OPENAI_KEY, REVIEWER_MODEL, HEADER_EXTRACTOR_MODEL
docker compose up --buildThen open http://localhost:8080. The backend API is at http://localhost:8000, GROBID at http://localhost:8070.
Note: the GROBID image is large and its first request after startup is slow while it loads its models — this is expected.