Building an AI Content Summarizer

My solution: A Python script that automatically processes URLs, generates summaries using local AI, and organizes them in Obsidian. No apis, no subscriptions, no cloud storage.

Why Build This? My 3 Motivations

1. The Quick Capture Bottleneck
   I kept saving interesting links to a “read later” file that became a graveyard of unprocessed content.
2. Context Switching Killer
   Manually summarizing YouTube videos, Instagram posts, and articles is tedious.
3. AI Assistant, Not Crutch
   I wanted an AI system that augments my thinking rather than replacing it.

The Tech Stack Breakdown

Core Components

• Obsidian as knowledge base (Why Obsidian?)
• Ollama for local LLM processing (using Deepseek-R1 8B model)
• transcribe-anything for video/audio processing
• Selenium for JS-heavy site handling

Key Features

1. Automatic URL detection and classification
2. Social media transcription pipeline
3. AI-assisted summarization with thought process preservation
4. Markdown formatting compatible with Zettelkasten

How It Works: A Technical Walkthrough

1. The Config Loader

class ConfigLoader:
"""Merge default and user configs"""
def load_config(self) -> Dict[str, Any]:
return {default_config, loaded_config}

Handles configuration with sensible defaults while allowing customization.

2. Content Processing Pipeline

def process_section(self, section_info) -> str:
if "instagram.com" in url:
return self.process_instagram(url)
elif "youtube.com" in url:
return self.process_youtube(url)
else:

return self.summarize_webpage(url)

Different processing paths for different content types while maintaining a unified interface.

3. AI Interaction

def clean_response(self, text: str) -> str:
"""Keep thought patterns visible but separate"""
return re.sub(r'<think>.?</think>', '', text, flags=re.DOTALL)

Preserves AI’s reasoning in special tags while keeping final output clean.

Why This Beats Manual Note-Taking

• Time
  Processing a YouTube video went from 5-10 minutes to 30 seconds
• Consistent Formatting
  All notes follow Zettelkasten principles automatically
• Local Processing
  No data leaves my machine (crucial for client work)
• Visible Thinking
  AI’s reasoning is captured but separated from final output

Challenges & Lessons Learned

1. Selenium Headaches
   Had to implement smart retries and explicit waits for JS-heavy sites:

WebDriverWait(driver, 10).until(
EC.presence_of_element_located((By.TAG_NAME, "article"))
)

1. Model Hallucinations
   Added strict prompt engineering to keep outputs focused:

prompt = f"Respond only with the summary and hashtags: {text[:2000]}"

1. Config Management
   Created a configuration loader that merges defaults with user overrides.