AIHOT

AI news monitoring and curation website built by Chinese AI media creator 卡兹克. URL: aihot.virxact.com

Overview

• Sources: 168 curated information sources (RSS, HTML scraping, APIs, paid data interfaces)
• Daily volume: ~563 articles captured, ~50% filtered out as non-AI-related
• Models used: DeepSeek V3.2 (pre-screening), DeepSeek V4 Pro (scoring)
• Free to use (public version has limited features vs. internal enterprise version)

Architecture

1. Source collection with T1/T1.5/T2 tiering
2. AI relevance pre-screening (DeepSeek V3.2)
3. 5-dimension scoring (DeepSeek V4 Pro)
4. Weight calculation via code formulas (model only scores, doesn’t decide curation)
5. Event clustering via embeddings
6. Daily auto-generated AI report (5 categories, 1-second generation)

Key Lesson

Iterated scoring strategy 11 times. Final approach: separate model scoring from code-based decision making. Model only does 5-dimension scoring; everything else (weighting, thresholds, curation) handled by deterministic code. Prompt reduced from 600 lines to 200 lines after this separation.