CatchUp.AI

How the Catch Up summery is being shown
How the articles are shown, and if there's any new update since you last read
Landing page
Landing page in Dark Mode
Article page in Dark mode

Inspiration

CatchUp.AI was born out of personal frustration: trying to keep up with the ongoing drama between Elon Musk and Donald Trump. Each time I returned to the news, I had to sift through dozens of articles, many of them repetitive or outdated, to figure out what was actually new. I realised there had to be a smarter way to track evolving stories, like how developers track versions in software. That simple idea became the foundation of CatchUp.AI.

What it does

CatchUp.AI uses AI agents to track news stories over time. It generates versioned summaries for each topic, and when a user returns, it highlights the exact updates since they last read, as well-written personalised news articles. The app also recommends trending or related topics based on user interests.

How we built it

We built an agentic pipeline with:

"NewsFetcherAgent": Pulls articles from the web, RSS, and APIs.
"TopicClusterer": Groups related articles using embeddings and vector DB.
"SummaryAgent": Creates versioned summaries using Gemini.
"DiffAgent": Writes new 'Catch Up' articles highlighting the differences between versions since the user's last read.
"CurationEngine": Personalises trending and relevant topic feeds.

All versioned articles and diff articles generated by the AI pipeline will power the app’s personalized content delivery. When a user views a topic, the system checks if a diff article (between their last-read version and the latest) already exists. If not, the app will generate it on the fly using the AI summarization engine, store it, and serve it, ensuring fast, relevant catch-ups for every user without redundancy.

I deployed the AI Workflow on GCP, leveraging Supabase for storage, Vertex AI for LLM operations, ElevenLabs for reading articles, and Netlify for the frontend & API hosting. Everything except the AI pipeline was built with 'Bolt.new', still generating the diff articles based on users' last seen versions were built with Bolt.new.

Challenges we ran into

Designing meaningful versions that are both lightweight and informative.
Handling long input contexts when summarising multiple articles.
Keeping user experience simple while working with complex AI pipelines.

Accomplishments that we're proud of

Built a system that mirrors how developers think about versions, but for news.
Achieved real-time diff summaries using Gemini with grounding.
Developed a scalable curation engine that feels personal, not generic.

What we learned

News consumption needs structure as much as coverage.
Users value “what changed” more than re-reading entire stories.
LLMs, when guided with structure (versioning, diffs), become incredibly efficient assistants.

What's next for CatchUp.AI

Personalized Interest Tags: Introduce an “Interest Areas” feature that allows users to personalise their reading experience. Each article will display a set of pre-defined tags such as “Economic Impact,” “Ethical Perspective,” “Political Context,” etc. When a user clicks on a tag, the article dynamically expands or adjusts to highlight that perspective, giving readers insights tailored to what they care about most.
Add deeper personalisation: fine-tuned feeds based on tone, topic, or source trust.
Launch browser extension to summarise the current page or alert to updates.
B2B API: Offer our version/diff summaries to newsrooms and analysts.
Multi-modal: Extend to podcasts, YouTube transcripts, and social media threads.

Limitations

We have limited the number of topics tracked by our AI Pipeline to keep the costs minimal.
AI Pipeline is running every 24 hours to keep the costs minimal.