OZGPT
Closing The Digital Divide with Intelligent Solutions
Project Showcase
Full-page screenshots of the live application
The Challenge
AI assistants are typically confined to web and mobile applications, limiting accessibility for users without smartphones or reliable internet access.
The objective was to launch Pakistan’s first telecom-integrated, multi-channel AI assistant for Zong Network subscribers, accessible via:
Web platform
Mobile app
SMS
USSD
IVR (voice interface)
This required solving several high-complexity challenges:
Bridging modern LLM infrastructure with legacy telecom protocols (SMS, USSD, IVR)
Managing LLM context window limitations at scale
Handling multilingual queries and real-time translations
Implementing sentiment-aware responses
Optimizing cost efficiency across millions of interactions
Ensuring uninterrupted availability under rapid user growth
Scaling conversational AI across low-bandwidth and non-smartphone channels was particularly complex, as these environments lack persistent sessions and rich UI context.
The Solution
A layered AI orchestration architecture was developed to support seamless interaction across all channels.
Key components included:
Hybrid AI Stack
Integrated multiple LLM APIs for conversational intelligence, paired with a dedicated web-search augmentation layer to improve factual accuracy while optimizing inference costs.
Context & Memory Management Engine
Designed a context compression and session handling layer to intelligently manage token limits across constrained channels such as SMS and USSD.
Sentiment & Intent Layer
Implemented real-time sentiment detection and intent classification to dynamically adapt responses and improve conversational quality.
Cost Governance Framework
Built a usage monitoring and cost-control layer to track API consumption, dynamically allocate workloads, and prevent service disruption during high traffic spikes.
Telecom Channel Abstraction Layer
Developed a unified routing system that translates conversational AI responses into formats compatible with SMS, USSD, IVR, web, and mobile environments.
The result was a scalable, telecom-grade AI ecosystem accessible to both smartphone and feature-phone users.
Tech Stack
The Results
The platform achieved significant adoption and operational milestones:
4.8M+ registered users within first 6 months
350K+ Daily Active Users (DAU) across all channels
1.5M+ monthly interactions processed
Enabled AI access to non-smartphone users for the first time at national scale
99.3% service uptime across channels
Reduced per-interaction inference cost by an estimated 38% through hybrid orchestration
4.7/5 average user rating on mobile platforms
The platform established a new benchmark for AI accessibility in emerging markets by bridging advanced language models with traditional telecom infrastructure.