pAIper research review console

pAIper is a venue-aware AI peer review console for research manuscripts. It reads a draft, scores it against a structured editorial rubric, checks citation integrity, evaluates fit for target venues, and turns the result into a prioritized revision plan.

The tool is designed for transportation, computer science, AI, computer vision, NLP, machine learning, robotics, and adjacent applied research manuscripts.

What pAIper does

pAIper is not a passage-level grammar checker. It is a manuscript readiness system that behaves more like a strict associate editor:

• Reviews the manuscript across contribution, related work, methods, statistical validity, results, reproducibility, ethics, writing quality, technical correctness, and claim integrity.
• Produces an editorial recommendation, overall score, evidence-backed findings, strengths, weaknesses, and revision priorities.
• Checks target-venue fit using curated venue profiles and submission expectations.
• Verifies citations against public scholarly indexes and flags missing, inconsistent, fabricated, or chimeric references.
• Runs a citation-only audit when users want reference verification without a full manuscript review.
• Supports cloud providers and an Ollama / Local option for offline private review.
• Exports results as Markdown, Word, PDF, and JSON.

Core capabilities

Capability	Description
Venue-aware review	Optional venue targeting for transportation journals, TRB, IEEE venues, CVPR, ACL, NeurIPS, ICLR, ICML, AAAI, KDD, SIGSPATIAL, CoRL, and related outlets.
Multi-pass review engine	Runs independent review passes so a weak model response cannot collapse the full report.
Citation verification	DOI-first checks followed by Crossref, Semantic Scholar, and OpenAlex lookups.
Citation-only mode	Verifies references without running the full review rubric. BibTeX uploads can be parsed directly without a model provider.
LaTeX bibliography support	Prompts for .bib or .bibtex files when citation checks are requested for LaTeX manuscripts.
Claim-integrity audit	Identifies unsupported claims, overreach, missing evidence, and internal consistency problems.
Export pipeline	Generates structured in-app reports plus DOCX, Markdown, PDF, and JSON outputs.
Provider flexibility	Supports OpenRouter, Gemini, OpenAI, Anthropic, Mistral, and Ollama-compatible local endpoints.

Supported inputs

• PDF
• DOCX
• LaTeX
• Markdown
• Plain text
• arXiv URLs
• BibTeX files for citation verification

PDF parsing uses PyMuPDF by default. OCR and parser behavior are controlled from the application sidebar.

Quick start

git clone https://github.com/pozapas/pAIper.git
cd pAIper
python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt
streamlit run app.py

On macOS or Linux:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
streamlit run app.py

Then open the local Streamlit URL shown in the terminal.

Basic workflow

1. Choose Full review or Citation check only.
2. Select a model provider in the sidebar.
3. Paste an API key for cloud providers, or choose Ollama / Local for a local endpoint.
4. For full review, choose review depth: Fast, Standard, or Deep.
5. Optionally select a target venue.
6. Upload a manuscript, paste an arXiv URL, and optionally attach a .bib or .bibtex file.
7. Run the review or citation check and inspect the verdict, findings, venue fit, citation checks, and action plan.
8. Export the result in the format needed for revision or collaboration.

Local models and privacy

pAIper can call Ollama through its native local HTTP API. For private offline review, run pAIper on the same machine as Ollama:

ollama pull llama3.1:8b
ollama serve
streamlit run app.py

Then select Ollama / Local and keep the endpoint as http://localhost:11434. pAIper reads the endpoint's installed model list and fills the Model dropdown automatically.

On https://paiper.streamlit.app/, localhost is the Streamlit Cloud server, not the user's device. The hosted app can use an Ollama-compatible endpoint only if that endpoint is reachable from the cloud. Users who do not want manuscripts to pass through an online system should run pAIper locally.

Review depth

Depth	Intended use	Review coverage
Fast	Early draft triage	Contribution, methods, results, and writing.
Standard	Default pre-submission check	Core editorial dimensions plus venue and citation options.
Deep	Final Q1-style readiness pass	Expanded technical correctness and claim-integrity review.

Repository layout

app.py                  Streamlit application shell
engine/                 UI-independent review engine
engine/exporters/       Markdown, Word, and PDF exporters
ui/                     Theme, icons, and reusable Streamlit components
ui/assets/              App imagery and local UI assets
data/venues/            Venue profiles and checklists
docs/assets/            README and documentation assets
.streamlit/config.toml  Streamlit runtime theme and upload settings
requirements.txt        Python dependencies

Citation verification

Citation verification is designed to be conservative. pAIper checks references through public scholarly services and reports uncertainty instead of inventing matches. It can help identify:

• Missing DOI or bibliographic metadata.
• Real papers cited with wrong authors or venue details.
• Chimeric citations where the title and metadata do not belong together.
• References that cannot be found in public indexes.

LaTeX manuscripts often contain citation keys such as \cite{smith2024} rather than full reference metadata. When citation checks are enabled for a .tex upload, pAIper asks for the matching .bib or .bibtex file.

In Citation check only mode:

• A BibTeX file can be checked directly without a model provider.
• A manuscript or arXiv URL without BibTeX can still be checked, but pAIper needs a model to extract structured references before querying Crossref, Semantic Scholar, and OpenAlex.
• The output focuses on references, verified matches, suspicious records, and export files.

Citation checks depend on external service availability and should be treated as decision support, not a replacement for manual reference review.

Model and cost notes

pAIper sends review prompts to the provider selected in the sidebar. Costs depend on the chosen model, review depth, manuscript length, and citation-verification settings. For quick testing, OpenRouter and Gemini free-tier models can be used where available.

The app keeps review passes modular so smaller or cheaper models can still complete useful reviews. Stronger models generally produce better evidence synthesis and more reliable revision priorities. Ollama costs no provider tokens, but local quality and speed depend on the model and hardware.

Privacy

• API keys entered in the sidebar are kept in the browser session.
• The app sends manuscript text to the selected model provider for review unless Ollama / Local is used in a local offline run.
• Citation verification may query Crossref, Semantic Scholar, and OpenAlex.
• On the hosted Streamlit app, uploaded manuscripts pass through Streamlit Cloud even when the selected model endpoint is external.
• Do not upload confidential or restricted manuscripts unless your chosen provider and deployment environment meet your data-governance requirements.

Deployment

pAIper is a Streamlit application. It can be deployed to Streamlit Community Cloud, Hugging Face Spaces, Render, Railway, or a private server that supports long-running Python requests.

Minimum deployment requirements:

• Python 3.10 or newer.
• Dependencies from requirements.txt.
• A secrets mechanism for provider API keys if you do not want users to paste keys in the UI.
• Enough request time for long manuscripts and deep reviews.

Development

Run a local syntax check:

python -m py_compile app.py ui\components.py ui\theme.py engine\engine.py

Run the app:

streamlit run app.py

Keep development artifacts out of commits. The repository intentionally excludes .codex, .claude, .agents, virtual environments, logs, local Streamlit secrets, caches, and generated output folders.