3arabeetak: Analysis of Architecture Design and Technical Implementation of a Multimodal AI Automotive Platform

The 3arabeetak project addresses pain points in used car transactions in emerging markets like Egypt—such as information asymmetry, difficulty in vehicle condition assessment, and opaque pricing—by building an end-to-end multimodal AI automotive service platform. It integrates technologies like Playwright parallel crawlers, YOLOv8/ViT four-stage visual assessment, machine learning price prediction, and a locally deployed Gemma-3 chatbot to form a complete intelligent service system, providing efficient solutions for both buyers and sellers.

In emerging markets like Egypt, used car transactions have long faced pain points such as information asymmetry, difficulty in vehicle condition assessment, and opaque pricing. Traditional processes rely on personal experience or intermediaries, which are time-consuming and prone to pitfalls. The 3arabeetak project targets this demand by building an end-to-end multimodal AI platform that integrates computer vision, natural language processing, and predictive machine learning. As a graduation project, it demonstrates the complete technical chain of modern full-stack AI applications: from data collection and intelligent analysis to interactive recommendations.

3arabeetak adopts a layered architecture with four core subsystems:

Data Collection Layer: Parallel crawlers based on Playwright handle dynamic JS pages to ensure data integrity and timeliness.

Visual Analysis Layer: Four-stage vehicle condition assessment: YOLOv8 object detection to locate key parts → ViT fine-grained feature extraction → multi-view fusion → damage classification, outputting a structured report.

Price Intelligence Layer: Machine learning models combine vehicle model, year, mileage, condition score, etc., to output EGP reference price ranges, reducing negotiation friction.

Interactive Intelligence Layer: A locally deployed Gemma-3 chatbot ensures response speed and privacy, supporting natural language search, comparison, and consultation.

YOLOv8+ViT Combination: YOLOv8 has strong real-time performance, suitable for vehicle photo streams; ViT's self-attention captures long-range dependencies and identifies subtle damages, forming a complement.

Playwright Parallelization: To deal with anti-crawling mechanisms, reasonable intervals and distributed scheduling balance efficiency and compliance; PostgreSQL's JSONB fields support semi-structured data storage and querying.

Local Deployment of Gemma-3: Under Egypt's network conditions, local deployment solves the problems of high cloud latency and cost. It can run on consumer-grade hardware, ensuring privacy and offline use.

Targeting Egypt's special car import policies, the platform has a built-in import tariff calculator. It integrates complex rules such as customs duties, consumption taxes, and value-added taxes. Users can input basic vehicle information to get an accurate estimate of the total landed cost, which is of great practical value to buyers of imported vehicles.

Insights from the 3arabeetak project:

1. Multimodal Fusion Trend: A single modality is difficult to solve complex problems; joint modeling of visual + text + structured data has become a standard.
2. Value of Edge Deployment: In specific regions/sensitive scenarios, local deployment has significant advantages in cost, latency, and privacy.
3. Domain Knowledge Engineering: The import calculator requires deep integration of industry rules, reflecting the evolution from general AI to domain-specific AI.
4. End-to-End Experience: Technical complexity is hidden behind a simple interface; the competitiveness of AI products lies in the efficiency of solving problems.

As an academic project, 3arabeetak's technology selection and architecture design have high engineering reference value, providing a reusable blueprint for AI applications in vertical fields. With the evolution of multimodal large model technology, similar intelligent service platforms will take root in more industries.