Ask Hassan an Arabic-speaking AI assistant with a realistic avatar and a governed knowledge base
A digital presenter that answers by voice and video, and only from approved sources.
- 54
- lip-synced clips
- 80%
- accuracy on 188 questions
- 69
- approved entries
- 0
- invented numbers in 21 voice answers
The problem
A nonprofit gets the same questions every day and cannot afford an assistant that invents a number or a fact about its services.
Overview
An Arabic-speaking assistant with a face and a voice. Visitors type or press the mic and get back a lip-synced video answer. It is the second project for the same client after the management system, and it is now embedded in their public website.
What I built
- 54 clips produced with a LatentSync pipeline on an hourly-rented GPU
- Three-layer question matching measured at 80% accuracy on 188 test questions
- A knowledge base of 69 approved entries drawn from 7 source files
- A visible “virtual character” badge on every surface, following SDAIA disclosure guidance
- Staff update Hassan’s facts from the management system, no developer needed
The engineering call
The hardest problem was the face disappearing mid-answer. The fix is a three-layer stage switched by audio level, with a threshold of 0.006 taken from the measured gap between speech and silence over 1,136 audio windows.
