ArabsStock AI Engineer 2023–2024

ArabsStock Assistant: 14 agents and AI search over millions of assets

Architected a 14-agent LangGraph assistant for ArabsStock and the semantic and visual search over millions of its photos, videos and vector graphics.

See the architecture

Role: Architecture and build (AI Engineer) Full-time, Remote
Company: ArabsStock an Arab stock-media platform
When: Apr 2023 – Dec 2024 1 yr 9 mos
Where: Remote (Dubai, UAE)
Stack: LangGraph
OpenAI API
Pinecone
Python
FastAPI
PostgreSQL
Docker

specialised agents in one LangGraph assistant: 14
vision and audio model services, each a FastAPI container: 5
of photos, videos and vector graphics behind semantic search: Millions
Apr 2023 – Dec 2024, full-time and remote: 21 months

Fig. 1 arabsstock-assistant · architecture

Figure 1 The ArabsStock Assistant routes platform requests through a LangGraph orchestrator to 14 specialised agents, backed by semantic search on Pinecone, containerized model-serving endpoints and media pipelines. Diagrams show the components named in public descriptions of the work. They are simplified, not complete system maps.

Orchestrator. LangGraph routes each request to the right specialised agent.
Specialised agents. 14 of them, covering semantic search, AI image generation, database operations, content moderation and integrated AI services.
Evaluation. LLM agent evaluation workflows monitor reliability, response accuracy and agent routing in production.
Semantic search. Automated batch ingestion processes and indexes millions of photos, videos and vectors with OpenAI embeddings in Pinecone.
AI services. Model-serving endpoints for image search, sound generation, instrument classification, BPM detection and audio watermarking, each a containerized FastAPI microservice.
Media pipelines. Production pipelines take raw video and audio through quality optimisation and metadata enrichment before distribution on the platform.

Text description of the diagram

Architecture diagram. Top: media pipelines take raw video and audio through quality optimisation and metadata enrichment, then deliver it to the ArabsStock platform (photos, videos, vectors and audio). Below the platform are three systems. Assistant: platform requests go to a LangGraph orchestrator, which routes them to 14 specialised agents grouped by capability: AI image generation, content moderation, semantic search, integrated AI services and database operations. Semantic search queries the Pinecone vector index; database operations use the platform database. An LLM agent evaluation workflow monitors response accuracy, agent routing and reliability. Indexing: platform assets go through automated batch ingestion and OpenAI embeddings into the Pinecone vector index, covering millions of assets. Model serving: five containerized FastAPI microservices serve image search, sound generation, instrument classification, BPM detection and audio watermarking to the platform.

On this page The problem

The problem

ArabsStock is an Arab library of millions of royalty-free photos, videos and vector graphics. The platform needed one assistant for very different jobs (finding assets, generating images, running database operations, moderating content) and search that finds assets by meaning and by visual similarity.

Approach

Built the semantic search pipeline: OpenAI embeddings into a Pinecone index, filled by automated batch ingestion across millions of assets.
Added visual-similarity search, so users can find visually similar assets across the library.
Served image search, sound generation, instrument classification, BPM detection and audio watermarking as containerised FastAPI microservices, and added sound search for a beta sound library.
Added agent evaluation workflows that track reliability, response accuracy and routing in production.

Architecture

Component view: a LangGraph orchestrator routes requests to specialised agents; a batch pipeline keeps the vector index current; model services run as separate containers; evaluation watches routing and accuracy. Figure 1

Users and apps: Platform users
Orchestration: LangGraph orchestrator
Agents: Semantic search
Image generation
Database operations
Content moderation
Other platform AI services
Models: OpenAI embeddings
Retrieval: Visual-similarity search
Data: Pinecone vector index
Services: Audio model services · Generation, search, instruments, BPM, watermarking
Pipelines: Batch ingestion · Millions of photos, videos, vector graphics
Quality: Agent evaluation
Connections: Platform users to LangGraph orchestrator
LangGraph orchestrator to Semantic search
LangGraph orchestrator to Image generation
LangGraph orchestrator to Database operations
LangGraph orchestrator to Content moderation
LangGraph orchestrator to Other platform AI services
Batch ingestion to OpenAI embeddings
OpenAI embeddings to Pinecone vector index
Agent evaluation to LangGraph orchestrator (routing, accuracy)

Outcome

A production assistant whose 14 agents cover search, image generation, data operations and moderation.
Millions of photos, videos and vector graphics indexed for semantic search, plus visual-similarity search across the library.
A new audio line for the platform: a beta sound library with generation and search.
Agent reliability, accuracy and routing measured in production rather than assumed.

Stack

Orchestration: LangGraph
Models: OpenAI APIvLLM
Retrieval: Pinecone
Data: PostgreSQL
Services: PythonFastAPI
Infrastructure: DockerNVIDIA A100

Where these facts come from. Everything on this page comes from my CV, my LinkedIn profile and my GitHub profile and repositories. Nothing is estimated: where no figure is public, the page describes what the system does.

Next case study

Wakeb Data2024–2026

A 6-agent bilingual chatbot on self-hosted Qwen

Architected a fully async 6-agent LangGraph chatbot in Arabic and English, and ran its Qwen models on vLLM across production A100 servers.

LangGraph
Qwen
vLLM
NVIDIA A100
FastAPI

Questions about this build?

Ask my AI about “ArabsStock Assistant: 14 agents”.

It answers from my CV and public profile, in English or Arabic.

Contact me or write to hatem@mohamedhatem.dev