ARA

Problem & Goal

ARA was born from a real need in RDFZ's research courses: frontier papers blocked by jargon and structure. The goal: an AI tool that lets non-specialists understand academic literature fast — and give back to the classroom.

Award: Excellence Project Award (1st place), Zhongguancun Academy 2025 Youth AI Camp.

Core Features

• AI document parsing: PDF / Word / Excel / images → structured Markdown + Mermaid mind-maps
• AIDIG figure workspace: auto-generated scholarly illustrations with split-screen editing
• Multi-model × multi-script engine: Gemini / DeepSeek, scripts extensible by admins
• Full user system: JWT auth, file management, task tracking, 24 admin endpoints
• Security: TOTP 2FA, IP allowlist, login lockout, audit logs, HTTPS

Multi-Agent Pipeline

MoE architecture with 4 agents in relay: Expert Manager classifies discipline (~100 tokens, fast model) → Discipline Expert writes a deep review (humanities / social / natural science / engineering routing) → Citation Checker verifies every claim against the source → Mermaid Creator builds the mind-map in parallel.

OpenRouter model routing sends simple tasks to light models and deep reasoning to strong ones, cutting API cost by ~40%.

Real Feedback

In a blind test, two high-school teachers compared three reviews of the same paper (ARA v1, ARA v2, raw DeepSeek). Both rated ARA significantly better in content quality and logic.

Outcomes

• v1.1 live on a self-managed Windows server with HTTPS
• 10,000+ LOC backend: user system, payments, multi-agent engine, admin panel
• Validated in blind tests by real teachers against raw model output
• 1st place Excellence Award, Zhongguancun Academy 2025 Youth AI Camp