Veritas

Inspiration

Researchers lose hours to busywork between tools, about 14 hours just to format a single paper. AI tools like NotebookLM can answer questions about papers, but nothing checks what you actually wrote. Software has linters that catch bugs before code ships, and we wanted the same thing for research papers.

What it does

Veritas is an IDE that finds the problems in your paper before reviewers do. It checks every claim against the paper it cites, and flags numbers that don’t match your data and sentences copied from a source. It turns your papers into a comparison table with every cell linked to the page it came from. When you’re done, it exports a submission-ready paper in IEEE, ACM or APA format.

How we built it

The frontend uses Next.js and the Monaco editor, the editor behind VS Code. The backend uses FastAPI and PostgreSQL with pgvector for searching papers. Google Gemini powers the citation audit, the literature matrix and the chat. Everything runs locally in Docker.

Challenges we ran into

Gemini’s free-tier limits stopped us partway through importing papers, so we built rate limiting, fallback across models, and automatic retries. Real-world PDFs were messy and broke our parser. Our first checks also raised false alarms, so we tested them against papers with planted mistakes until they caught every one without false alarms.

Accomplishments that we’re proud of

In testing, the citation audit caught a claim about BERT that was cited to the wrong paper and suggested the correct one. All four journal templates compile cleanly on the first try. Every AI judgment shows the passage and page it’s based on, so a researcher can check it in seconds.

What we learned

AI is most useful in research when it checks writing against evidence, not when it writes for you. Simple checks that don’t use AI catch a surprising number of real errors at no cost. Building on free tiers also taught us to design for failure from the start.

What’s next for Veritas

We want to link each reviewer response to the exact edit that addresses it, and add checklists for each venue’s submission rules. We’d also like an AI mock reviewer that critiques a draft against a conference’s real review criteria before you submit.

Built With

• python
• typescript