Amine AYARI

Principal-level individual contributor with 15+ years of experience designing and scaling production-grade AI systems across NLP, computer vision, 3D, embedded, and cloud platforms.

3 minute read

Published on April 16, 2025

NewsChrono

newschrono.com is a French short-form news platform I designed and developed end-to-end — from ideation and product design to backend architecture and production deployment.

The core idea behind NewsChrono is simple:

help users stay informed through concise, readable news summaries, while still enabling discovery of related topics and context.

Behind this simplicity sits a production-grade architecture that combines LLMs, semantic embeddings, serverless infrastructure, and scalable backend services. This article walks through the product vision, technical choices, and system design.

1. Product Vision: Short-Form, Not Shallow

Traditional news platforms overwhelm users with:

  • long articles,
  • redundant coverage,
  • noisy feeds.

NewsChrono focuses on:

  • short, high-signal summaries,
  • fast consumption,
  • semantic navigation across related stories.

The goal is not to replace journalism, but to optimize access to information for users who want to understand what matters quickly.

2. LLM-Based News Summarization

At the heart of NewsChrono is automatic article summarization.

Each full-length news article is processed using:

  • OpenAI LLM APIs,
  • prompt-engineered summarization logic,
  • consistent output formatting.

The summarization pipeline:

  1. Ingests the raw article
  2. Applies an LLM-based summarization prompt
  3. Produces a concise, readable summary
  4. Stores both raw and summarized content

This approach ensures:

  • consistent tone and length,
  • language clarity,
  • scalability across sources and topics.

Summarization is treated as a backend service, not a frontend feature.

Short-form content is more powerful when paired with contextual discovery.

To achieve this, each article is represented using:

  • LLM-generated embeddings,
  • a dense vector capturing semantic meaning.

Embedding Pipeline

For each article:

  1. Generate an embedding using an LLM embedding model
  2. Store the embedding vector in the database
  3. Use it as a semantic descriptor of the article

When displaying an article, NewsChrono retrieves:

  • the 4 most related articles,
  • based on cosine similarity between embeddings.

This is implemented using:

  • Firebase find_nearest
  • cosine distance as the similarity metric

The result is a recommendation system based on meaning, not keywords or categories.

4. Backend Services: Systemd-Managed Workers

Both:

  • article summarization,
  • and embedding generation + similarity indexing

are handled by a dedicated backend service, running as a systemd-managed process.

This design choice provides:

  • isolation from the web frontend,
  • reliable background processing,
  • controlled retries and monitoring.

The backend service is responsible for:

  • calling OpenAI APIs,
  • managing embeddings,
  • updating Firebase records,
  • triggering similarity indexing.

This separation keeps the system robust and maintainable.

5. Web Application Architecture

The NewsChrono web application is built as a:

  • FastAPI application
  • fully Dockerized
  • deployed on Google Cloud Run

Why FastAPI

  • clean API design,
  • async-friendly,
  • easy integration with backend services.

Why Cloud Run

  • serverless scaling,
  • zero infrastructure management,
  • cost efficiency,
  • fast iteration cycles.

Cloud Run allows the platform to scale automatically with traffic, without over-provisioning resources.

6. Firebase for Data Management

Firebase is used as the main data backend, handling:

  • article storage,
  • summaries,
  • embeddings,
  • metadata,
  • similarity queries.

Key benefits:

  • managed infrastructure,
  • tight integration with Cloud Run,
  • native support for vector similarity queries.

Firebase acts as both:

  • a traditional database,
  • and a lightweight semantic index.

7. End-to-End Data Flow

Putting it all together:

  1. News article is ingested
  2. Backend service summarizes it using an LLM
  3. Embedding is generated and stored
  4. Article is indexed for similarity search
  5. Web app fetches summary + related articles
  6. User consumes concise news with semantic context

Each component is loosely coupled but clearly defined.

8. From Ideation to Production

NewsChrono is not a prototype — it is a complete product lifecycle:

  • idea and positioning,
  • system design,
  • LLM integration,
  • backend services,
  • cloud deployment,
  • production operation.

Building it end-to-end required combining:

  • product thinking,
  • AI engineering,
  • backend architecture,
  • and cloud infrastructure.

Closing Thoughts

NewsChrono demonstrates how LLMs can be productized responsibly, not just experimented with.

By combining:

  • summarization,
  • semantic embeddings,
  • scalable backend services,
  • and serverless deployment,

the platform turns raw news streams into structured, digestible information.

This project reflects a broader belief:

AI is most powerful when it quietly improves how people access and understand information.

You May Also Enjoy