GivePap reads your PDFs, your Excel files, your R scripts, and returns a structured scientific paper draft β Introduction, Methods, Results, Discussion, Conclusion + References with real DOIs. We don't write the paper for you: we give you back 3 weeks a month.
A Markdown draft structured section by section, with citations verified via DOI lookup and honest [NEEDS DETAIL] tags where the experimental data is actually missing.
[NEEDS DETAIL] where the data is missing β it doesn't invent the gap.ChatGPT with one prompt generates a page that looks like a paper. No coherence between sections. Hallucinated DOIs. Invented numbers. GivePap uses 6 specialized agents that cooperate β and check each other.
The standard objection against AI on papers is: "it invents the numbers." That's true for plain ChatGPT. It's not true here β because a second, independent agent reopens the file and re-verifies every figure in the draft against the source cell.
| lactation_stage | F_value | p_value |
|---|---|---|
| Early | 3.31 | < 0.001 |
| Mid | 2.84 | 0.003 |
| Late | 1.97 | 0.041 |
"ANOVA revealed a significant effect of lactation stage on miR-148a expression (F(9,452)=3.31, p<0.001), with early-lactation cows showing 2.3-fold higher levels than late-lactation cows."
No plausible-looking numbers. No hallucinated DOIs. No citations made to look real.
A second agent reopens the Excel files and re-verifies every number in the draft. Mismatch β visible flag.
Every [Author Year] passes through PubMed E-utilities, CrossRef and OpenAlex. If it's not there, it's flagged.
When the experimental data is actually missing (e.g. spectrometer model), the tag stays. No silent invention.
Every citation passes through a chain of external lookups. The final format is pluggable β the same pipeline produces JDS or Nature or IEEE without rewriting the paper.
Fallback chain: if PubMed doesn't find a DOI, try CrossRef. If CrossRef also fails, OpenAlex. If none β explicit flag on the citation, it doesn't get invented.
Style selectable at the run level. Change tier or change target journal, regenerate without rewriting β the formatter is separate from the writer.
Four cases where we tell you directly to look elsewhere. We'd rather lose a customer than sell them a tool they don't need.
Python 3.12 / FastAPI backend with 15,600 lines of code, 24 REST endpoints, 7 DB tables, Multi-agent pipeline in testing with partner researchers. Deploy via Docker Compose, data self-hosted in the EU.
Primary model for the higher tiers. Sonnet 4.6 for the fast draft. Tier-aware routing.
Data inside your perimeter. Anthropic zero-retention opt-out active on the pipeline.
PostgreSQL 16 + pgvector extension for semantic embedding on uploaded PDFs.
6 specialized agents, orchestration via Redis 7 + RQ worker.
JDS, Vancouver, APA, Nature, Cell, ACS, ACM, IEEE, PhysRev, Elsevier Harvard, Harvard.
Three output formats. Markdown is the source, DOCX and PDF generated on the fly.
Export the files at any time. The pipeline produces a draft, not a hosting service.
We're running real tests with partner researchers. Public numbers will land after this phase.
Single-file deploy. Next.js 15 + React 19 frontend, FastAPI backend, Postgres, Redis.
Private beta open to 10 partner researchers β free for 3 months in exchange for structured feedback. Are you building a paper right now? Let's show you β we run the pipeline on your dataset, you see the draft, you decide if you need it.