Mistral Worldwide Hackathon: LUMINAE

🌍 Live App • 📺 YouTube Demo • ✨ Features • 🏗️ Architecture • 🧠 AI Agents

🌍 Live Application: https://luminae.qzz.io/
📊 Analysis Dashboard: https://luminae.qzz.io/dashboard
📺 YouTube Demo (2 min): Watch on YouTube

Luminae is an autonomous research analysis platform that transforms academic papers into actionable intelligence. Using Mistral AI's multimodal capabilities, it performs deep critique, identifies methodological flaws, proposes novel experiments, and generates grant proposals — all within a strict $15 API budget through intelligent caching and cost optimization.

Upload any document (or paste a public URL) and Luminae's 5 AI agents deliver:

Luminae is a deliberate showcase of Mistral's latest multimodal ecosystem, with models explicitly chosen for their unmatched cost-to-performance ratios:

• mistral-ocr-latest (The Digester): Handles the heavy lifting of extracting highly complex academic PDFs. We use Mistral OCR to perfectly preserve the markdown structure, headers, and dense tables of scientific literature without losing the document's flow.
• ministral-3b-2512 (The Visualizer): Powers our Vision Analysis parallel stream. As figures are parsed from the paper by the OCR engine, this ultra-fast 3B model interprets charts, graphs, and diagrams into text descriptions, feeding vital visual context back into the reasoning pipeline.
• ministral-8b-2512 (The Reasoning Engine): The core "brain" powering all 5 of our diverse reasoning agents in parallel. We rely strictly on Mistral's native Structured Output constraint (response_format={"type": "json_object"}). Every single agent is locked into outputting highly complex JSON schemas ensuring robust data parsing.

Note: Models were aggressively optimized to ministral-8b-2512 coupled with smart Redis caching to guarantee high volume paper processing without breaking the $15.00 hackathon budget limit.

• Frontend: Next.js 16.1 (App Router), TypeScript, Tailwind CSS, shadcn/ui, TanStack Query, Zustand
• Backend: FastAPI, Python 3.11, Pydantic, HTTPX
• Database & Auth: Supabase (PostgreSQL, Row Level Security)
• Cache & Rate Limiting: Redis (Upstash)
• AI Layer: Mistral AI APIs with Budget Protection Middleware

Steps 2a/2b/2c run in parallel via asyncio.gather:

Step 1: OCR  →  Step 2a: Vision Analysis     ┐
                Step 2b: Methodology Critique  ├→  Step 3: Experiments  →  Step 4: Synthesis  →  Step 5: Grant
                Step 2c: Dataset Audit        ┘

/luminae
├── apps/
│   ├── api/                   # FastAPI Backend (Python)
│   └── web/                   # Next.js Frontend (React)
├── packages/
│   ├── mistral-client/        # Shared Mistral Client wrapper
│   ├── types/                 # Shared TypeScript interfaces
│   └── ui/                    # Shared UI Components
├── schema.sql                 # Supabase Database Schema
├── turbo.json                 # Turborepo configuration
└── vercel.json                # Vercel Deployment configuration

Rather than relying on a single generic reasoning prompt, Luminae operates a highly concurrent multi-agent swarm via asyncio.gather. Each of the 5 specialized AI personas has a singular focus, preventing context dilution and generating extremely deep, expert-level analysis:

1. 🔬 The Methodology Critic: Evaluates the paper’s experimental design. It actively hunts for confounding variables, selection biases, and statistical power issues, providing a severity-rated critique and an overall reproducibility score.
2. 📊 The Dataset Auditor: Analyzes the underlying data structures. It flags missing data handling strategies (MAR/MCAR), hidden demographic biases, and ethical or IRB oversights in the sample groups.
3. 🧪 The Experiment Designer: Steps beyond summary and enters pure creation. It invents 3 entirely novel follow-up experiments designed to address the specific flaws found by the Critic and Auditor, delivering strict hypotheses, methods, feasibility scores, and budget estimates.
4. 🔗 The Synthesis Agent: A high-level supervisor that cross-references the findings from the Critic, Auditor, and Experimentalist to resolve conflicts and output 3 counter-intuitive "Key Insights" that would surprise the original authors.
5. 📝 The Grant Generator: Translates the proposed experiments and synthesis into a structured, NSF-style funding proposal outline, complete with specific aims, timelines, and a research strategy.

Each agent is strictly constrained to JSON schemas via Mistral's response_format={"type": "json_object"} mode, ensuring a type-safe, parseable, and hallucination-free data pipeline.

1. Create a new Web Service in Render
2. Connect to this GitHub repository
3. Configure:
   • Root Directory: apps/api
   • Build Command: pip install -r requirements.txt
   • Start Command: uvicorn src.main:app --host 0.0.0.0 --port $PORT
4. Environment Variables:

1. Import this repository in Vercel
2. Settings:
   • Framework: Next.js
   • Root Directory: apps/web
   • ✅ Include source files outside of the Root Directory
3. Environment Variables:

Run /schema.sql in your Supabase SQL Editor to create papers, analyses, and usage_logs tables with RLS policies.

3-layer cost protection ensures we never exceed the $15 budget:

1. Pre-flight Estimation — Approximates tokens before API call. Rejects requests > $0.50
2. Global Circuit Breaker — Redis-backed countdown. Halts if remaining < $2.00
3. Usage Logging — Exact costs logged to Supabase for audit

Cost per paper: ~$0.05 → ~300 papers on $15 budget

Built with 🧠 by DevRay Labs © 2026