You can access the GitHub link for this product here:

KrishivSIyer / deskreject-gaurd

DeskReject Guard

Local, fully offline pre-flight auditor for academic manuscripts that catches visual, anonymity, and formatting desk-reject flaws before submission.

Team

Team Name: Tomato Chutney

Member	Contribution
Krishiv S Iyer (Team Leader)	Vision client, Ollama worker pool, geometry/color checks, and image anonymity
Aditya Vineeth	Streamlit UI, visual overlays, synthetic test samples, and documentation
Karthik M G	Text checks (anonymity, mentions, sequencing, statements) and LaTeX patch generator
Sibi Chakravarthi	Core PDF parsing, ingest pipeline, report finalization, and evaluation suite

Problem Statement

The Problem

Academic manuscripts submitted to top conferences and journals must adhere to strict formatting, anonymity, and submission guidelines. Authors frequently encounter desk-rejections—rejection prior to peer review—due to preventable formatting and anonymity violations: author identities leaked in document metadata or hyperlinked repositories, university crests and lab logos embedded in system architecture diagrams, figure citations appearing out of order, missing mandatory ethics or reproducibility statements, unreadable tiny chart text, or captions…

View on GitHub

Sorry for the spelling error in the repo 😭. We noticed last minute.

Imagine spending months on groundbreaking research, compiling your LaTeX file, and submitting it, only to get tossed out in 48 hours because a reviewer couldn't read the font on your scatter plot. Between 15% and 30% of academic submissions get desk-rejected instantly due to completely preventable visual and formatting rules.

Standard proofreading tools check grammar, but they are totally blind to the visual layout of a PDF. They don't know if your chart labels are too small or if you accidentally left your university logo on a hardware photo during a double-blind review.

For Hacktoberfest Hack Day Coimbatore 2026, we built DeskReject Guard to fix this. It’s a fully local, multimodal pre-flight inspector that visually audits your manuscript against specific journal guidelines and actually hands you the LaTeX code to fix the errors.

What It Actually Does

Instead of staring at a 20-page journal guideline PDF, you just drag and drop your draft into our web UI. Here is what the engine handles under the hood:

• The Vision Parity Check: It visually counts the sub-panels in a multi-part figure and cross-references them against your caption text. If Figure 3 has a visual panel (d) but your text only explains (a) through (c), it catches it.

• Double-Blind Deanonymization Scrubber: It scrubs text for unblinded GitHub links and possessive self-citations. More importantly, it uses visual OCR on system diagrams and hardware photos to flag hidden lab watermarks or university crests.

• Micro-Typography Guard: It calculates the bounding boxes of text embedded inside your charts. It warns you if your axis labels will shrink below the mandatory 6pt minimum when forced into a standard 3.5-inch publication column.

• Colorblind Accessibility Auditor: It runs your heatmaps and multi-line plots through simulated colorblindness filters (protanopia, deuteranopia). If your red/green data points are indistinguishable, it suggests accessible Matplotlib hex schemes (like Viridis or Cividis).

• Compliance Sweeper: It vector-searches your back matter to ensure mandatory disclosures—like Data Availability, Code Sharing, and IRB approvals—are present and in the correct template section.

• LaTeX Patching: Instead of vague advice, the app hands you a copy-pasteable LaTeX diff to fix the issue. It will automatically convert \documentclass[final] to anonymous review variants and adjust your figure environments.

The Hackathon Architecture (100% Offline)

Because the hackathon mandated zero cloud dependency and local AI execution, we couldn't just throw API calls at OpenAI. We had to build a distributed edge-computing cluster to process the heavy multimodal tasks within 5 hours.

• The Brain (Primary Node): We ran Streamlit and FastAPI on a Lenovo Legion (RTX 5070 Ti, 12GB VRAM). This handled the heavy text reasoning using a customized gemma4:12b-it-qat model. When a violation was found, we routed the parameters to qwen2.5-coder:14b to generate the LaTeX diffs.

• The Eyes (Edge Worker Nodes): We distributed the actual computer vision tasks to three teammate laptops (RTX 4050, 6GB VRAM) acting as parallel microservices. These ran gemma4:e4b-it-qat natively to process image crops, evaluate typography using OpenCV, and return JSON reports over a local hotspot.

The final output is an interactive dashboard that overlays red bounding boxes directly onto your PDF where the formatting fails, alongside a drawer of instant LaTeX fixes. We built DeskReject Guard so researchers can stop fighting formatting rules and get back to actual science.

Comments

[ai_adam]

Desk rejection là nỗi đau thực sự — nhất là khi reviewer chỉ lướt qua 30 giây rồi bấm reject vì thiếu checklist item nhỏ nhặt. Idea guard tại local/CI trước khi push lên remote rất practical, đỡ tốn cycle review của team.

Thấy bạn nhắc GitHub link, có thể share thêm chút về architecture không? Ví dụ: rule engine chạy như thế nào (static analysis, regex, LLM-based?), config format ra sao để team custom cho repo mình? Còn edge case: false positive rate như thế nào khi apply cho codebase legacy đang refactor dần?

Nếu có benchmark so sánh với Husky + commitlint hoặc custom GitHub Actions thì còn hay hơn — giúp team quyết định adopt nhanh hơn PS: the tool I meant is on labagent .tech