Digital Sharjah Assistant - Multi-Agent Government AI Assistant

Context

Digital Sharjah is the Emirate of Sharjah’s unified government super-app and portal, on iOS, Android and web, in Arabic and English. It launched in 2021 and had 50+ services by 2022. It brings together UAE Pass sign-in, multiple digital payment methods, and integrated services from government entities and sub-platforms such as Aqari (real estate), Mazraaty (agriculture), Tarkhees (business licensing), SEWA, Sharjah Police and Civil Defence.

The platform is architected for ~1M users, 10K concurrent sessions and 10K transactions per second. It ships at a near-weekly cadence - ~50 web and mobile releases in 6 months and 160+ roadmap features over 2 years - and its support operation handles ~40K requests a year at ~97% resolution.

The Digital Sharjah Assistant is the AI layer on top of that platform. I am its Technical Manager.

What is live

A multi-agent LLM assistant: an intent-routing orchestrator, 7+ specialised service agents and a retrieval-augmented (RAG) knowledge agent. It is live on web, iOS and Android, in Arabic and English, and covers parking, traffic fines, utilities, telecom, flight status and business licensing.

I am also Technical Manager for the assistant’s real-time voice experience: streaming voice conversation and read-aloud on mobile.

Roadmap and transactional skills

I led the assistant’s roadmap and its move from answering questions to completing transactions:

• Traffic fines - defined a 13-week, 5-phase plan for inquiry and payment across 2 government data sources, with duplicate removal and payment deep links.
• Smart parking - finds the 5 nearest open car parks, with ticket inquiry and payment.
• Bilingual conversation design across Arabic and English.

Quality and stabilisation

I led a quality and stabilisation program covering issues across 4 delivery streams: bilingual regression testing, resilience of the retrieval and integration layers, and end-to-end observability.

Responsible-AI transparency

I designed the assistant’s transparency UX. Users see live progress steps while the assistant runs a task. Model reasoning, prompts and personal data are never exposed.

AI-to-human handover

I defined the handover from the assistant to the government contact centre, with 6 delivery tracks covering UX, UI and front end: escalation detection, live-agent routing, agent identity and state, attachments and end-of-chat flows.

Next: a government-wide agentic layer

I authored the business case for next-generation agentic orchestration, which turns the assistant into a government-wide AI layer. It is still in planning. It covers:

• A multi-agent architecture with MCP (Model Context Protocol)-based service onboarding.
• An AI governance framework.
• A build-vs-buy agent-platform proof of concept.

What it targets:

• Scale - 30+ services from 5+ entities, grounded in 85+ knowledge sources.
• Onboarding - 70%+ of new services onboarded via MCP, in under 5 days each.
• Quality - over 90% intent accuracy, at least 95% task completion, under 2 s response time, 100+ concurrent sessions, CSAT of 4.0/5 or higher, and near-zero hallucination.
• Expected impact - 40%+ fewer in-person visits.

Agent-to-agent integration

I am Technical Manager for agent-to-agent integration with a partner platform’s AI agent. It established a reusable pattern: a new partner agent is added by configuration, a prompt and a specialist sub-agent. It is designed for a ~1-second response time.

Real-time knowledge pipeline

I am Technical Manager for a real-time knowledge pipeline that feeds the assistant. It uses change data capture, event streaming and a bilingual RAG index. It makes zero changes to the source system, rolls out in 5 phases and targets 15-minute data freshness. The Architecture Review Board approved it.

AI-powered matching service

I am Technical Manager for an AI-powered matching service for a government entity:

• Multimodal (image + text), cross-lingual matching.
• Officer-in-the-loop approval.
• Privacy by design: on-premises OCR, hashed identifiers, and no personal data sent to cloud models.

Pilot targets: a match confirmed within 48 hours, officer review in 2 minutes or less, 80%+ of found items handed in through the service, and CSAT of 4.2/5 or higher.

Why it matters

Most government AI projects start as point solutions. The Digital Sharjah Assistant is being built the other way round: one shared assistant that routes intents, coordinates service agents, and gives government services a standard way to connect to AI. Transparency, human handover and privacy are part of the design from the start.

Themes

This project sits where three topics I often write about meet: multi-agent system design, what it actually takes to run agentic AI, and how a Technical Program Manager delivers AI in regulated environments with many stakeholders. The writing section has more on each.