Luminae
Five-agent research-paper analysis built on Mistral models
A Mistral Worldwide Hackathon entry that reads a research paper with Mistral OCR and runs five specialised agents over it: a methodology critique, a dataset audit, experiment suggestions and a grant outline, then a synthesis. Outputs are structured and cost-controlled, and they are meant as decision support.
At a glance
- Status
- Live web app
- Type of work
- Hackathon entry
- Context
- Mistral Worldwide Hackathon
- Categories
- Applied AI & Agents · Web & SaaS
- Source code
- Public repository
Technology
Problem
Reviewing a paper’s methodology, data and possible next experiments takes expert time. Students and small research teams benefit from a structured first pass before they spend that time.
Who it’s for
Researchers, students and small research teams.
What I built
- Document ingestion with Mistral OCR, including figures.
- Five specialised agents running over the parsed paper: Methodology Critic, Dataset Auditor, Experiment Designer, Grant Generator and a Synthesis Agent that combines their findings.
- Structured model outputs so every agent returns predictable, renderable results.
- FastAPI backend, Next.js frontend, Supabase PostgreSQL and Redis (Upstash) caching.
- Budget middleware that tracks and limits API usage.
Architecture
- IngestUpload → Mistral OCR and figure understanding.
- AnalyseFour specialist agents run over the parsed document.
- SynthesiseSynthesis Agent combines the specialist outputs.
- ControlStructured outputs, Redis caching and budget middleware.
- Store & presentSupabase PostgreSQL; Next.js interface.
Engineering challenges
Orchestrating several agents
Specialist agents work in parallel and a synthesis step reconciles their findings into one report.
Predictable output
Structured outputs keep each agent’s result in a consistent shape that the interface can render reliably.
Staying within a hackathon budget
Caching avoids re-analysing the same paper, and budget middleware caps spend.
Results
| Measure | Value | Evidence |
|---|---|---|
| Estimated cost | ≈ $0.05 per paper within a $15 hackathon budget | README estimate, not independently measured |
Limitations
- Outputs are AI-generated decision support. They are not verified research conclusions and can contain errors.
- Hackathon entry; no award or ranking is claimed.