← Projects

My SDAIA Co-op Project

Deployed

A fine-tuned Arabic model that turns any tech article into a newsletter-ready summary.

Overview

For my final project during the SDAIA internship, I built an internal tool for the newsletter team that helps them quickly and accurately summarize any tech article. They can drop in a URL or raw text, and the tool generates an Arabic summary written in the same style as the SDAIA newsletter, with the English version and the original article displayed side by side so every line can be matched back to its exact source.

Key Features

  • ❖ Pulls and extracts article content from URLs or direct text.
  • ❖ Uses RAG to ensure factual grounding and prevent model hallucinations.
  • ❖ Highlights every generated sentence and shows its precise source in the original article.
  • ❖ Fine-tuned GPT-4o-mini Arabic model trained on 800+ newsletter articles.

The Challenge

Delivering an Arabic summary that matches SDAIA’s editorial quality and helps the team speed up production, while keeping every sentence fully source-backed.

The Solution

The system first extracts the article, then routes it through an AI agent to verify quality and extraction accuracy. After that, it flows through a RAG step to generate the initial outputs. These outputs are merged into a dedicated summarization prompt to refine and validate the content. Once the summary is ready, it’s translated into Arabic using the fine-tuned Arabic model. Each sentence is then aligned with its original location in the source article using a larger model. Finally, everything is presented in a dual-pane interface with highlights, editing tools, copy options, text resizing, and more.

Gallery

Stack

Next.js TypeScript LangChain RAG Fine-tuned LLM model ExtractorAPI

Project Links

View project ↗ View code ⌘