Building a Sovereign AI native Media OS for Arabic Media

Build a Sovereign AI Media OS: Engineer Bluejay-an in-house, model-agnostic agentic engine built from first principles on PydanticAI specifically designed to power Arabic and English newsrooms across SRMG.

Unify Fragmented Workflows: Replace isolated drafting, research, and multimedia tools with specialized AI studio apps and no-code agent workflows that handle end-to-end media operations from research to publishing.

Automate Proactive Intelligence: Deploy background scheduler agents to monitor trends overnight, qualify breaking news, and deliver structured morning briefs to editors.

Embed responsible-AI governance: by design, with explainability, guardrails, mandatory human approval, and a full audit trail.

Maintain Dynamic Cost Control: Implement per-run token and cost tracking across 400+ LLM models to maintain transparent budget management.

Operational Scale & Velocity: Generated 13,000+ Al content summaries, processed 10,000+ news feeds, and produced 1,000+ content rewrites with 300+ editor refinements, cutting breaking-news turnaround from hours to minutes.

Ecosystem Integration: Live integration with G.O.A.T app and Billboard Arabia CMS, powering 100% of G.O.A.T app content summaries via Bluejay APIs. 

100% Governed Adoption: Achieved zero ungoverned publication paths; every output passes mandatory human-review gates and maintains full traceability to sources and models. 

Data Sovereignty & Cost Control: Safeguarded embargoed material within SRMG infrastructure using on-premise models while maintaining transparent, per-desk cost attribution across 400+ LLM models.

Agentic Harness : Built on top of PydanticAI libraries with integrated concepts and capabilities for Tooling, MCP integrations, DAG based Multi agent workflows, Agent skills, Compaction, RAG based Knowledge base

LLM providers: OpenAI, Anthropic, Google Gemini, xAI, Groq, Amazon Bedrock, DeepSeek, Moonshot, Xiaomi MiMo, OpenRouter, Ollama

Data and retrieval: Milvus vector database (RAG), AWS VectorStore, [SQLite / PostgreSQL], [Redis]

Media pipeline: text-to-speech (Gemini, OpenAI, AWS Polly, Kokoro), Speech to Text (Google Gemini, Open AI) , Video Models (Gemini, OpenAI, Seedance)

Automation and integration: scheduled automations, REST API and webhook callers, Playwright browser automation

Governance and observability: in-house run ledger, usage and cost hub, exportable explainability traces

SRMG enterprise systems: CMS, Analytics and marketing platforms

Reimagining an AI-native Media OS required building an executive-grade personal agent for creators and Editors-in-Chief spanning research to publishing. Key challenges included:

Fragmented Workflows: Unifying isolated research, drafting, and multimedia tools into a single platform backed by enterprise governance.

Multi-Publication Scaling: Supporting multi-lingual (Arabic/English) brand guidelines and house styles without codebase fragmentation.

Data Sovereignty & Security: Protecting unpublished stories on SRMG-controlled infrastructure via custom on-premise model deployments.

Editorial Accountability: Enforcing human-in-the-loop approvals with full audit trails across sources, models, and decisions.

Cost & Reliability: Controlling LLM consumption across providers while ensuring long-running processes survive system restarts.

Change Management: Shifting editorial teams from ad-hoc prompting to governed, structured workflows through targeted enablement.

Bluejay is SRMG's in-house AI native Media OS for newsrooms. It replaces open-ended chatbots with a home screen of specialized agents and studio apps that run editorial work as repeatable, governed pipelines. Editors fill a form or pick a saved configuration; agents research, draft and package the story; the run pauses whenever tone, audience or legal sensitivity is unclear; and nothing reaches readers without an editor's approval. Every run is logged with its model, sources, tool calls, tokens and cost. It is natively integrated with news wires for a seamless content creation workflow end to end. The core features are

Harness + Loop Engineering : Pydantic AI-native runtime with DAG flow with integrated capabilities for Multi agent orchestration, tool calling, MCP and Skills integration.

Media Apps Ecosystem : Integrated AI native apps tailored for media professionals. Podcast creator, Image Editor, Social video creator , Newsletter Creator, Social Video creator etc.

No Code Agent Builder : A no code graphical agent and multi agent workflow builder for content creators to build and automate their own agentic workflows

Explainability and Traceability : Every agentic input, reasoning, tool calls, outputs logged as exportable and auditable traces to govern and control performance, cost and any adversarial agentic patterns.

Model and Cost Controls : Integrated model library to help keep costs in control and use the best cheapest model for any given task.

Proactive Workflows. An editor configured scheduler runs in the background looking for the breaking stories, and handles it to the editors second brain to qualify and produce the first draft. It comes back to the editor with the rationale for selection and a first draft of all content formats – reducing the time from hours to minutes and making a reactive process to a time sensitive proactive push.

Deep Research Agents. Research flows pull sources into a governed library and build long-form drafts with citation blocks, pausing for naming and redaction decisions. Sensitive material can run on on-premise models. The creators now have an intelligent researcher to drill down and extract nuances for any story with valid citations.

Agentic Knowledgebase: Integrated Vector DB for using any unstructured data as a RAG based context for the agents and integrated datasets module to bring in two way read-write functionality for agents to work with structured data,

Automations for scale. Overnight automations set the scene for the day through monitoring of trends and topics of importance. The agents keep a daily intelligence brief ready before the morning sets in.

Content Repurpose at scale. An approved article becomes social posts, a podcast episode with AI voices, a faceless short video and a slide summary, each inheriting the approved source.

Kingdom of Saudi Arabia