DataUp

DataUp is a data infrastructure and AI data collection company specializing in high-quality training datasets, annotation pipelines, and agent management for AI startups and research institutions across North Africa and the Middle East.

I joined DataUp as a Full-Stack Software Developer to strengthen their core engineering platform. My mandate focused on two key technical objectives:

1. Designing and building a centralized AI Agents Management Dashboard to orchestrate, monitor, and evaluate autonomous data collection jobs in real time.
2. Resolving severe rendering and data-processing bottlenecks in CVAT (Computer Vision Annotation Tool) to support complex Arabic right-to-left (RTL) text extraction, document parsing, and bounding-box annotations.

1. AI Agents Management Dashboard

Architected and built a full-stack operational control room that allowed data teams and clients to deploy, monitor, and debug autonomous web-crawling and data-collection agents:

• Real-Time Agent Telemetry: Managed live status streams, task queues, memory logs, and error reporting across dozens of concurrent scraper and agent tasks.
• Granular Task Control: Implemented interactive controls to pause, resume, re-point, or adjust rate-limiting parameters on active agents without requiring terminal access.
• Data Quality Verification: Built pipeline interfaces for annotators to inspect raw scraped payloads before passing them to validation queues.

2. CVAT Core Modification for Arabic Right-to-Left (RTL) Support

CVAT's open-source frontend and backend had fundamental bugs when handling Arabic text annotation for OCR and document vision models:

• The Technical Challenge: Bidirectional (BiDi) text algorithms in standard HTML Canvas elements caused Arabic characters to render reversed, disconnected, or visually misaligned with bounding-box labels. Additionally, string manipulation in the backend API stripped diacritics and corrupted character positions during bounding-box spatial indexing.
• The Solution: Engineered a core open-source modification patch. Modified CVAT's HTML5 Canvas rendering engine using custom BiDi re-ordering logic (bidi-js integration) to ensure Arabic script maintained proper ligature connection and right-to-left alignment inside active canvas bounding boxes.
• Backend Normalization: Extended CVAT's Django backend models and serializers with UTF-8 Unicode normalization logic, ensuring spatial text offsets matched DB records exactly during export.