Sam's space

I recently built a small and opinionated AI news aggregator site (ai.samu.space) to help filter through the noise and find meaningful AI-related news and research. The project took about 3-4 hours to put together, leveraging FastAPI, React, news APIs and OpenAI’s API.

# Architecture

The application consists of three main components:

• A FastAPI backend that fetches and processes news articles, storing them in sqlite
• A React frontend for displaying the news in a clean interface
• An LLM service that helps filter and summarize the content

# News Processing Pipeline

The most interesting part of the system is how it processes incoming news articles:

1. Articles are fetched from multiple sources (NewsAPI, academic papers)
2. Each article goes through content filtering:
   • Blocked terms filter (to avoid celebrity drama and ignore people I don’t want to hear about)
   • Similarity check to prevent duplicates
   • AI-relevance check using LLM
3. Valid articles are stored in a sqlite database
4. A summary is generated for the latest batch of news

The LLM service acts as a smart filter, determining if an article is truly AI-related rather than just containing buzzwords. This helps maintain signal-to-noise ratio in the feed.

# Frontend Experience

The frontend is intentionally minimal, focusing on readability and quick access to information:

• Clean card-based layout for articles
• Tabs for switching between news and academic content
• AI-generated synthesis of latest developments
• Mobile-responsive design

# Technical Decisions

Some interesting technical choices I made:

# Background Processing

Rather than making users wait, article fetching happens in the background using FastAPI’s BackgroundTasks. When a user requests news, they immediately get cached results while a refresh is triggered if needed.

# Caching Strategy

Articles are cached for 6 hours to balance freshness with API rate limits. The cache invalidation is handled automatically when users request news.

# Content Deduplication

To avoid duplicate content, the service uses a combination of title matching and content similarity checking. This handles cases where the same news appears on different sites with slightly different titles.

# Learnings

Building this highlighted some interesting challenges:

• LLM APIs can be slow - had to carefully balance thoroughness of filtering with response times
• News APIs have varying quality - some return duplicates or irrelevant content requiring additional filtering
• Background tasks need careful error handling to prevent failed tasks from affecting the main application
• Summarization requires finding the right prompt balance between conciseness and informativeness

The project was a fun exercise in putting together different services (news APIs, LLMs, frontend) into a cohesive application. While simple, it serves its purpose of helping me stay updated on meaningful AI developments without getting lost in the noise.