llm-tracker Project Introduction

llm-tracker is an open-source, self-updating dashboard project that automatically monitors, transcribes, and analyzes content from 12 popular AI/LLM YouTube channels. Using dual-layer transcription, LLM-powered intelligent analysis, and semantic topic association, it helps users quickly grasp the latest trends in the AI field and cross-channel consensus, transforming scattered video content into a structured, searchable knowledge base.

Project Background

In the era of information explosion, the AI field sees a large number of technical videos released daily, covering topics like large model architectures and inference optimization. Manually tracking the latest developments has become increasingly difficult. llm-tracker aims to address this pain point by building a fully automated system for in-depth analysis and integration of video content.

Core Function Architecture

Dual-Layer Transcription System

• Layer 1: Prioritize using YouTube auto-generated subtitles (obtained via yt-dlp tool), fast and low-cost;
• Layer 2: Cloudflare Whisper API as backup to ensure content acquisition for videos without auto-subtitles.

Intelligent Content Analysis Layer

Driven by Hermes AI Agent, it performs multi-dimensional analysis of videos: content summary, specific topic extraction (562 sub-topics), key insights, creator stance, technical depth rating, and highlight quotes.

Cross-Channel Topic Association

Uses Qwen3-Embedding-0.6B to generate semantic vectors, discovers related topics across different channels via cosine similarity, and reveals industry consensus and multi-perspective views.

Technical Implementation and Channel Ecosystem

Data Processing Pipeline

Runs daily at UTC 17:00 and 20:00: Video discovery → Transcription → Keyword extraction → LLM-enhanced analysis → Semantic embedding → Association calculation → Data update.

Topic Classification System

• Basic Categories: 20 (GPT, LLaMA, Claude, RAG, etc.);
• Specific Sub-topics: 562 (e.g., BPE tokenization, KV cache optimization).

Monitored Channels

Covers 12 AI-related channels, including Andrej Karpathy, 3Blue1Brown, Two Minute Papers, etc., spanning dimensions like theoretical research, engineering practice, and educational popularization.

Application Scenarios and Interface Experience

Dashboard Interface

Uses static HTML + D3.js: Word cloud view (popular topics), topic relationship graph (semantic association), filters (channel/date/topic), transcription viewer (full content).

Potential Application Scenarios

• Researchers: Quickly understand industry hotspots and consensus;
• Developers: Track practices of specific technologies (e.g., RAG, Agents);
• Creators: Find content inspiration;
• Learners: Filter high-value videos;
• Investors: Identify technical trends and market opportunities.

Limitations and Future Directions

Current Limitations

• Language limitation: Mainly processes English content;
• Real-time performance: Updates twice daily, cannot meet minute-level needs;
• In-depth analysis: Highly technical content still requires manual supplementation.

Future Directions

• Support multi-language and more channels;
• Introduce real-time notification mechanism;
• Add user-customized queries and alerts;
• Link to knowledge bases like arXiv and GitHub.

Project Summary

llm-tracker combines mature tools (yt-dlp, Whisper, embedding models) with an innovative architecture to transform unstructured videos into a queryable knowledge base. It provides a powerful platform for AI field trackers, saving time and revealing cross-channel trends. It also demonstrates a cost-controllable and privacy-friendly automated system design, serving as a reference for similar applications.