Arabic AI Newsroom Assistant with a 3-Agent Agentic Loop

The Problem

Arabic newsrooms face a dual challenge:

Speed pressure – Breaking news demands fast content turnaround
Quality Control – AI-generated Arabic often lacks grammar, regional tone, and editorial standards.
Accountability – No audit trail when AI content goes wrong

Most AI writing tools give you one draft and call it done.
Editors either accept flawed output or rewrite from scratch.

My Solution

I built a 3-agent agentic pipeline where AI critiques and refines its own work before presenting options to the editor.

The Pipeline

Input: "ارتفاع أسعار النفط"
         ↓
🤖 AGENT 1: GENERATOR
   Creates Draft V1: 3 headline options, body, SEO, tags
         ↓
🔍 AGENT 2: CRITIC
   Reviews V1 section-by-section in Arabic
   Scores each section, flags issues
         ↓
✨ AGENT 3: REFINER
   Takes V1 + critique → produces improved Draft V2
         ↓
📊 EDITOR VIEW
   Side-by-side comparison: V1 vs V2
   Editor picks, edits, or overrides
         ↓
💾 AUDIT LOG
   Tracks: which version selected, what was edited, override %

Key Features

Two Generation Modes
- Simple Mode
- Agentic Mode

Arabic – First Design
- Right-to-left interface
- Arabic prompts and critique feedback
- Designed for Gulf newsroom standards

Human-in-the-loop Governance
- Editor always has final say
- Override tracking: system logs when editors change AI output
- Override percentage calculated and stored

Full Audit Trail

Every generation stores:

Original Input
Draft V1 (Generator output)
Critique (Critic feedback with scores)
Draft V2 (Refiner output)
Selected version (1 or 2)
Editor’s final body (if edited)
Override delta and percentage
Timing Metrics (gen_time_ms, critique_time_ms, refine_time_ms)

Archietecture

Evaluation Framework

AI Output is unpredictable. Every generation could fail silently. Editors catching mistakes is expensive.
I built a 20 case eval suite that validates LLM ouput against the same rules the production system enforces.

The Eval Set

Check	Test Cases
Headlines	5
Tags	5
Body Lenght	5
Slug	5
SEO description	5

Each case is a real editorial – style Arabic topic like:

اجتماع وزراء الخارجية الخليجيين لبحث التعاون الأمني المشترك

ارتفاع أسعار النفط العالمية وتأثيره على موازنات دول الخليج

Structural Checks

Rule	Validation
Headlines	Exactly 3
Tags	Exactly 5
Body Length	350-650 words
Slug	Arabic-only with hyphens
SEO description	<160 characters

Running Evals

node server/evals/run-evals.js http://localhost:5001

Output:

[PASS] #1 (Politics) score=8/10
[PASS] #2 (Politics) score=7/10
[FAIL] #3 (Politics)
        - body word count 312 outside [350, 650]
...
18/20 passed (90.0%)

Exit code 0 = all pass. Integrates with CI.

Why This Matters

Most AI writing tools ship without evaluation. When output fails, editors catch it – or worse, readers do.
This eval framework:

Catches structural failures before they reach the editor
Enables CI/CD – blocks deploys if evals regress
Documents Quality – 90% pass rate is auditable proof the system works

Tech Stack:

Frontend: Vanilla JS, Tailwind CSS
Backend: Node.js, Express
Database: SQLite
AI: Google Gemini 2.5 Flash (architected to swap to Jais Arabic LLM)

Why this Matters for Saudi Newsrooms:

Challenge	How this Solves it
AI makes grammar errors	Critic agent catches issues before editor sees them
No visibility into AI decisions	Full audit trail for every generation
Editors waste time fixing AI	V2 is already improved; editor reviews, not rewrites
Compliance concerns	Override tracking proves human oversight

Screenshots:

➡️ Mode Toggle

➡️ Agent Progress

➡️ Critique Panel

➡️ Side – by – Side Comparison

➡️ Demo Video

What I learned

Agentic loops need structure – Without clear handoffs between agents, output degrades
Arabic critique is harder – The critic prompt required careful tuning for proper Arabic Feedback
Audit trails build trust – Editors adopted faster when they could see exactly what the AI changed

Built With

Node.js + Express
SQLite
Google Gemini 2.5 Flash
Vanilla JS + Tailwind CSS
Figma (architecture diagrams)

Part of myAI PM portfolio demonstrating agentic AI patterns for enterprise content workflows.