How 100+ candidates become 20 curated papers — visualized Dijkstra-style
interactive
idle · press R to run
✓
Stage 1 · Source
🌐 OpenAlex API
5 research queries × 30 papers
Input: 5 search queries targeting
Mamba, SSL/MIM, Segmentation, Shape, Foundation Output:~100–150 candidates Filter: English only, has DOI
✓
Stage 2 · Filter
📋 Rule Engine
deterministic · zero-cost
Rules:
• Last 3 months only
• Deduplication (DOI-based)
• Remove previously collected Output:~50–80 papers
✓
Stage 3 · Embed
🧠 e5-small
intfloat/multilingual-e5-small · 384d
Input: title + abstract Method: Multi-Anchor cosine similarity
(5 weighted anchors: Mamba 0.4, Seg 0.3, MIM 0.2,
Shape 0.05, Foundation 0.05) + Venue bonus + Citation bonus Output:Top 30 by phase1 score Cache: embeddings saved to disk (84KB)
✓
Stage 4 · Score
🤖 DeepSeek API
deepseek-chat · LLM relevance scoring
Input: Top 30 papers (title + venue) Method: Batch prompt — rate each
paper 1–10 for medical image segmentation relevance Blend: 70% DeepSeek + 30% phase1 Output:Top 20 final papers
✓
Stage 5 · Deploy
🚀 Publish
HTML → VPS via SCP
Generate: Dark-themed HTML table Deploy: SCP to betsy.monster Schedule: Cron monthly (1st, 09:00 JST) Priority: nice 19 + ionice idle