I built TexGuardian after spending yet another deadline night fighting LaTeX formatting instead of focusing on research. Every conference submission, the same ritual: figure overflows, citation format issues, TODO markers left in text, hallucinated references from ChatGPT, forgotten anonymization. Hours wasted on mechanical formatting when you should be sleeping or refining ideas.
TexGuardian is a CLI that treats paper preparation like code review. It reads your entire .tex and .bib files, understands LaTeX structure and venue requirements, then generates reviewable unified diff patches for every issue it finds. No blind rewrites, no mysterious changes — every edit is shown as a diff you approve before it touches your files.
The `/review full` command runs a 7-step autonomous pipeline:
1. Compile with latexmk (proper error reporting)
2. Verify (figures, citations, TODOs, page limits, custom regex rules)
3. Fix issues with LLM-generated patches
4. Validate citations against CrossRef and Semantic Scholar APIs (catches hallucinated or outdated references)
5. Analyze figures (width overflows, placement, captions)
6. Analyze tables (booktabs compliance, column overflow)
7. Visual polish — renders PDF to images, sends to vision model, catches overlapping figures, bad spacing, margin violations that text-only analysis can't see
Key design decisions:
- Checkpoint system with instant rollback — every modification creates a restore point
- Unified diff patches only, never direct file writes — makes LLM edits auditable and reversible
- Async citation validation with concurrent API calls to CrossRef and Semantic Scholar
- Vision model convergence loop for PDF polish — iterates render → analyze → patch until quality stabilizes
- Natural language + slash commands — mix "/anonymize" with "fix the figure on line 303"
- Pluggable LLM backends (AWS Bedrock, OpenRouter) — default is Claude Opus 4.5 but supports any model
Works with 14 conference templates out of the box: NeurIPS, ICML, ICLR, AAAI, CVPR, ECCV, ACL, EMNLP, NAACL, COLING, CHI, KDD. Custom venue rules via regex patterns in paper_spec.md.
What started as a personal tool became something I thought the research community might find useful. If you've ever debugged \columnwidth calculations at 2 AM or validated 50 citations manually, this is for you.
pip install texguardian
Happy to answer questions about the architecture, LLM integration patterns, or take feature requests.
I built TexGuardian after spending yet another deadline night fighting LaTeX formatting instead of focusing on research. Every conference submission, the same ritual: figure overflows, citation format issues, TODO markers left in text, hallucinated references from ChatGPT, forgotten anonymization. Hours wasted on mechanical formatting when you should be sleeping or refining ideas.
TexGuardian is a CLI that treats paper preparation like code review. It reads your entire .tex and .bib files, understands LaTeX structure and venue requirements, then generates reviewable unified diff patches for every issue it finds. No blind rewrites, no mysterious changes — every edit is shown as a diff you approve before it touches your files.
The `/review full` command runs a 7-step autonomous pipeline: 1. Compile with latexmk (proper error reporting) 2. Verify (figures, citations, TODOs, page limits, custom regex rules) 3. Fix issues with LLM-generated patches 4. Validate citations against CrossRef and Semantic Scholar APIs (catches hallucinated or outdated references) 5. Analyze figures (width overflows, placement, captions) 6. Analyze tables (booktabs compliance, column overflow) 7. Visual polish — renders PDF to images, sends to vision model, catches overlapping figures, bad spacing, margin violations that text-only analysis can't see
Key design decisions: - Checkpoint system with instant rollback — every modification creates a restore point - Unified diff patches only, never direct file writes — makes LLM edits auditable and reversible - Async citation validation with concurrent API calls to CrossRef and Semantic Scholar - Vision model convergence loop for PDF polish — iterates render → analyze → patch until quality stabilizes - Natural language + slash commands — mix "/anonymize" with "fix the figure on line 303" - Pluggable LLM backends (AWS Bedrock, OpenRouter) — default is Claude Opus 4.5 but supports any model
Works with 14 conference templates out of the box: NeurIPS, ICML, ICLR, AAAI, CVPR, ECCV, ACL, EMNLP, NAACL, COLING, CHI, KDD. Custom venue rules via regex patterns in paper_spec.md.
What started as a personal tool became something I thought the research community might find useful. If you've ever debugged \columnwidth calculations at 2 AM or validated 50 citations manually, this is for you.
pip install texguardian
Happy to answer questions about the architecture, LLM integration patterns, or take feature requests.