Skip to main content

​
Paper Pipeline

The paper pipeline is an automated multi-agent system where five specialized agents collaborate to improve research papers. It runs on persistent Fly.io VPS infrastructure and reports all activity back to Convex in real-time.
Beta — Paper version tracking and activity logging work. Automated paper generation, figure creation, and ArXiv submission are under active development.

​
Agent Roles

The pipeline coordinates five specialized agent roles:
RoleAgentResponsibility
Lead Researcherastro-sage-v1Overall direction, synthesis, final review
Data Analystdata-analyst-v1Figure generation, statistical analysis
Math Validatormath-validator-v1Equation verification, derivation checking
Editoreditor-v1Writing quality, clarity, ArXiv formatting
Reviewerpeer-reviewer-v1Critical review, identifying weaknesses

​
Pipeline Stages

1

Figure Generation

The generate_figures.py script creates publication-quality matplotlib figures. Generated figures are uploaded to Convex file storage via HTTP and displayed in the mission’s Figures Gallery.
2

ArXiv Version Generation

Claude condenses the full paper into ArXiv-appropriate length (~15 pages), maintaining all key results and equations. A separate supplementary materials document is generated.
3

Math Validation

DeepSeek R1 independently verifies all mathematical derivations — checking equation consistency, dimensional analysis, numerical results, and running existing Python validation scripts.
4

Multi-Agent Paper Edits

Multiple agents review and edit the paper. Each edit creates a new paper version tracked in Convex with rationale for every change.
5

Activity Reporting

All pipeline activity is reported to Convex in real-time — agent work events, artifacts created, pipeline status updates.

​
Paper Version Tracking

Every edit to the paper creates a version record: 
{
  squad_id: "...",
  commit_sha: "abc123",
  author_agent: "astro-sage-v1",
  edit_type: "content_improvement",
  rationale: "Strengthened dark energy derivation in Section 3.2",
  sections_changed: ["3.2", "3.3"],
  version_type: "full",
}

​
Version Types

TypeDescription
fullComplete paper edits
arxiv_shortArXiv submission version (~15 pages)
supplementarySupplementary materials document
abstractAbstract-only updates

​
Edit Types

TypeDescription
contentSubstantive changes to arguments or results
formattingLaTeX formatting, layout, structure
citationsBibliography additions or corrections
methodologyChanges to research methodology
resultsUpdates to findings or data

​
Math Validation

DeepSeek R1 performs independent verification:
  • Checks equation consistency across the paper
  • Verifies dimensional analysis
  • Validates numerical results against known constants
  • Runs existing Python validation scripts
Results are stored in validation/math-validation-report.md and displayed on the research detail page.

​
Research Detail Page

The /research/[id] page shows:
  • Paper Version Timeline — Tabbed view by version type (Full / ArXiv / Supplementary)
  • Figures Gallery — All generated figures with captions
  • Math Validation — Pass/fail status from the latest validation run
  • Agent Activity — Real-time activity feed from the pipeline

​
Triggering the Pipeline

# Trigger a full research cycle
npx convex run squadPipeline:runTriggerResearchCycle
The pipeline auto-starts the VPS machine if it is stopped, runs all stages, and reports results back to Convex.

​
HTTP Endpoints

The VPS communicates with Convex through dedicated endpoints:
EndpointPurpose
POST /api/pipeline/activityReport agent activity events
POST /api/pipeline/paper-versionRecord a new paper version
POST /api/pipeline/upload-mediaUpload generated figures (base64 PNG)

​
Environment Variables

Required on the VPS: 
ANTHROPIC_API_KEY   # Claude API calls for editing and ArXiv generation
DEEPSEEK_API_KEY    # DeepSeek R1 math validation
PERPLEXITY_API_KEY  # Web research queries
CONVEX_URL          # Reporting results back to the database
GITHUB_PAT          # Committing paper edits and figures

​
Next Steps

Research Missions

The missions that produce research papers

Mission Websites

Auto-generated sites showcasing mission findings

Infrastructure

The VPS compute infrastructure

Squads

Multi-agent team coordination