The problem
The answer to an Arabic question often exists, but it is forty minutes into one of a hundred thousand videos, and no search box can reach it. Arabic makes it harder: spelling varies, diacritics come and go, and transcripts are noisy.
What we built
A pipeline for Arabic speech at scale
Transcription, Arabic normalization, AI classification and embeddings, run over a library of more than 131,000 videos in Python.
Hybrid search
Meaning-based vector search and full-text keyword search in one PostgreSQL database with pgvector, plus query expansion by a language model so a casual question reaches formal source language.
Answers you can check
Answers are retrieved from the source and linked back to the exact video, with Quran verse correction on top. Nothing is invented by the model.
A full right-to-left product
Laravel 12 with React 19 and Inertia, fully right-to-left, with shareable answer links.
Screens
How we worked
This is a studio product, designed and built in-house. It is where we test the retrieval and Arabic language work that shows up later in client projects.
Results
videos indexed and searchable
Transcribed, normalized and embedded.
Every answer links to the moment in the video it came from
Retrieval first, so answers can be checked against the source.
Fully right-to-left Arabic interface with shareable answers
What's next
The retrieval and Arabic normalization work here is the basis of the AI features we offer to clients.
