1 minute read

Tech news is fragmented across dozens of sources — RSS feeds, aggregators, social media, research feeds. Keeping up means scanning hundreds of articles daily, most of which are rewritten press releases or low-substance hype. We wanted a service that does the filtering for us: pull from everywhere, surface what actually matters, skip the noise.

It’s live at news.llm-works.ai.

What It Does

The service runs a pipeline every hour:

  1. Collect — pulls articles from 50+ RSS feeds and aggregators covering AI, programming, finance, crypto, science, and security
  2. Summarize & rate — each article gets a 2-3 sentence AI summary and a 1-5 star relevance score based on significance, novelty, and substance
  3. Cluster — related articles are grouped into thematic digests using semantic similarity
  4. Daily summary — top stories condensed into a short daily digest

Every item links back to its original source. Summaries are AI-generated, not pulled from the source.

Personas

Not everyone reads the news the same way. The service offers 10 personas — developer, founder, investor, researcher, executive, designer, marketer, student, journalist, lawyer — each filtering and ranking the same article pool differently based on what matters to that role.

A developer’s feed surfaces technical releases, benchmarks, and open-source launches. A founder’s feed highlights market moves, competitive signals, and funding rounds. Same source pool, different signal — the articles that matter to you float to the top.

Runs Local

All AI tasks — summarization, rating, clustering, and digest generation — run on a local Qwen 2.5 14B model. No article data is sent to third-party AI services.

The pipeline processes 300-500 articles per day on a single GPU. Running local keeps costs fixed and latency predictable — no rate limits, no per-token billing, no third-party data sharing.

Built on Our Stack

llm-news is built on our open-source stack: llm-saia for structured LLM operations (summarization, rating, and clustering all use typed verbs with output guards), llm-infer for model-agnostic inference, and appinfra for service lifecycle. The pipeline swaps between local and remote models without code changes.

Try It

Pick a persona that matches how you consume news and browse today’s digest at news.llm-works.ai.

Feedback welcome at @serendip_ml.

Updated: