Out of frustration of keeping up with AI news, I built hackobar. It fetches the AI realted news from mmuliple sources such as HN, arxiv, github trending repos, huggingface, many ai subreddits, twitter handles that cover AI, Tech news outelts, major ai lab blogs and of course the numerous AI news letters that are growing by day.

GitHub trending repos isn't "news" but as an engineer it matters. you can't read every arXiv paper or every HuggingFace drop. HN has the best discussions, Twitter has some, newsletters have some. I used to check 5 places and still felt behind.

Building this was fun. I had to face many technical challenges, even though I have used Claude Code to help build I think the decisions I took to mitigate those challenges are interesting and wanted to share with this crowd.

1 Fetching: I had to fetch different sources at different rate because fetching some on regular bases didn't make sense. for example, fetching AI labs blogs every few hours does not make sense but the HN front page ir reddit to get a sense what trending does

2 Filtering: Each articles fetched was filtered for specifc top level keywords, then Gemma 4 26B batch classifies each item was AI relavant or not.

3 Deduplication: I deduped using 3 layer pipeline. URL normalization, Jaccard on word tokens and Gemma semantic matching. The same arXiv paper can arrive from HuggingFace or HN, and might show up in three newsletters the same day. When a duplicate is found, the original news items cross-platform score gets bumped instead of the new item being silently dropped.

4 Scoring: Each item is scored based on the engagement, llm signal, cross platform signal, recency decay. And additionally each source has a weight component which keeps research from getting buried by social noise.

5 Summarization: For all the filtered and deduped, scored articles I use claude to generate a single line title and a 50 word summary. I have ensured the prompt to make hype free summary which concentrates on architecture, the benchmarks and why it matters to indivisuals.

Finally it surfaces 25 top/new cards in a swipeable feed for easier navigation, you can still read the orginal article if you want.

Initally I built this to work entirely on cloudflare free tier. Keeping every external HTTP call and every Supabase query under 50 subrequest per invocation limit. Eventually i had to go for paid worker as the cpu 10ms limit was bit challenging with improved deduplication.

Stack: Next.js on Cloudflare Pages, Hono API worker, Supabase, Anthropic Claude, Gemma via Cloudflare Workers AI.

Recently, I added another lens that filters the whole feed and highlights how this AI news impacts if you are builder, researcher, founder, invester, policy maker and healthcare worker. Honestly not sure if it's useful or gimmicky. Curious what you think.

hackobar.com | no paywall, no login.

Would love your feedback or thoughts.