A Head of AI Engineering who has shipped it.
Agentic AI, LLMs, RAG and MCP are easy to put on a slide. Below is each one with the production system behind it: what I built or led, where it runs, and the numbers it hits.
For hiring managers in 30 seconds
- Most recent role: Head of Engineering (AI), Etera AI — G42-powered, Aug 2024 – Sep 2026
- Now: Founder & CTO, Atrix — six products shipped in 2026
- AI in production: Multi-agent LLM systems, hybrid RAG, 5 MCP servers, real-time voice agent
- Leadership: 13+ years; teams of up to 32 across Europe, MENA and India
- Languages: TypeScript, Python, Rust, Go
- Based: Dubai — open to the Middle East, Europe and remote
Agentic AI & multi-agent systems
Agents that run inside a real product, under real permissions, with real latency budgets. Not demos.
- Etera AI (G42-powered): led the multi-agent LLM layer (LangChain, LangGraph, MCP) — hierarchical agent graphs with scoped memory across Restaurants, Experiences, Hotels and Flights
- The orchestration ran on 21+ Bun/ElysiaJS microservices on a Kafka event backbone on EKS, with Grafana/Loki observability
- Atrix: an in-app agent over the whole workspace (CRM, projects, hiring, HR, billing), with per-call permission checks
- This site: an AI twin with tool use that answers questions, qualifies leads and books calls
RAG, embeddings & retrieval
Retrieval is where most LLM products fail quietly. I build hybrid retrieval that holds up on real data.
- Etera AI: built a high-throughput embedding service (BAAI bge models, pgvector + Redis KNN) behind one hybrid-retrieval pipeline for four booking verticals
- Vector-augmented retrieval per vertical, feeding the agent graphs rather than a single global index
- Atrix Mail: PostgreSQL + pgvector semantic search across a multi-tenant mailbox store
- Written up in production detail: hybrid RAG with Postgres full-text + pgvector + reciprocal rank fusion
MCP servers & AI integrations
Five Model Context Protocol servers in production, from public read-only to OAuth 2.1 with scoped tokens.
- Atrix: remote MCP server with OAuth 2.1 and 49 scoped, revocable grants (write implies read), listed in the official MCP Registry. One connection reaches three sister products
- Atrix Mail: a 41-tool MCP server published on npm
- Atrix Analytics: an AI analyst exposed over MCP
- Etera AI: federated MCP behind the multi-agent layer
- This website: a public, read-only MCP server at /api/mcp you can add to Claude in one command
Real-time voice AI agents
Voice is the hardest latency budget in AI. I've built the whole stack, down to the SIP layer.
- VoDialer AI: eight Rust services on FreeSWITCH and Kamailio with a streaming speech-to-text → LLM → text-to-speech cascade (AssemblyAI, DeepSeek, ElevenLabs/Fish)
- ~1.2 s real reply gap on live phone calls; designed for 100+ concurrent conversations
- Searchie AI (CTO, acquired by oceain.ai): WebRTC media server (Janus, Kurento) with Amazon Rekognition emotion analysis
LLM engineering & reliability
Choosing, routing, testing and paying for models like an engineer, not a fan.
- Multi-provider in production: OpenAI, Anthropic, DeepSeek and OpenRouter, chosen per task on latency, cost and quality
- Model swaps decided by scenario test suites run on real conversation transcripts, not benchmarks
- Guardrails as code: permission-scoped tools, human confirmation before anything sends, pays or deletes, and tool output treated as data, not instructions
AI platform & data infrastructure
The unglamorous part that decides whether AI survives production: events, data, observability, cost.
- Event-driven backbones: Kafka (MSK) on EKS at Etera AI; Kafka on AKS at AllStars Digital
- Atrix Analytics: Rust + ClickHouse ingesting 42.7k events/s on one CPU at 66 ms p99
- Kubernetes, Terraform, GitOps and observability since Dubai Municipality (Istio, Argo CD, Prometheus, Grafana)
How I run an AI engineering team
Evals before features
No model, prompt or retrieval change ships without a test set built from real conversations. If we can't measure it, we're guessing in production.
Retrieval before fine-tuning
Most quality problems are context problems. Hybrid search, reranking and better chunking fix more than a fine-tune, at a fraction of the cost and lock-in.
Budgets for latency and cost
Every agent gets a latency budget and a cost-per-task number on day one. Voice taught me this: 1.2 seconds is a product decision, not an accident.
Agents act with scoped permissions
Narrow, revocable scopes; human confirmation before anything sends, pays or deletes; tool output treated as data. Security is an architecture choice, not a review step.
Small senior teams, AI-assisted delivery
One engineer with the right tooling now ships what a squad shipped in 2022. I hire for judgment, pair humans with agents, and keep the review bar high.
The leader still writes code
I shipped six products in 2026 myself. That keeps estimates honest, reviews sharp and architecture grounded in what actually works this quarter.
What I'm open to
Head / VP of AI Engineering
Own the AI platform and the team building it: agents, retrieval, evals, model strategy, and getting LLM features to production reliably.
CTO of an AI-native company
Full technical ownership — architecture, hiring, delivery and security — for a product where AI is the core, not a feature.
Fractional Chief AI Officer
For companies not ready for a full-time hire: AI strategy, build-vs-buy, vendor and model choices, and first production agents.
Questions hiring managers ask
What does a Head of AI Engineering do?
Owns how a company builds AI into its products: the team, the platform (models, retrieval, agents, evals, observability), the model and vendor strategy, and the path from prototype to reliable production. Unlike a research lead, the job is shipping — latency, cost, quality and security of AI features real users depend on.
What practical AI engineering experience does Moeid have?
At Etera AI (G42-powered) he led a federated-MCP, multi-agent LLM layer (LangChain, LangGraph) across four booking verticals, on 21+ microservices with a dedicated embedding service (BAAI bge, pgvector + Redis KNN) and hybrid retrieval. In 2026 he built a real-time voice agent (VoDialer AI, ~1.2 s reply gap on live phone calls) and five MCP servers in production, including Atrix's OAuth 2.1 server listed in the official MCP Registry.
Head of AI Engineering or CTO — which is Moeid a fit for?
Both. He is strongest as Head of AI Engineering in a product company that is serious about agents and LLMs, or as CTO of an AI-native company where the AI platform is the product. He has held CTO, Head of Engineering and VP Engineering titles and led teams of up to 32.
Which AI stack does he work with?
LangChain, LangGraph, MCP, the Vercel AI SDK; OpenAI, Anthropic, DeepSeek and OpenRouter models; pgvector, Redis KNN and Postgres full-text for retrieval; AssemblyAI, ElevenLabs, FreeSWITCH and Kamailio for voice — on TypeScript, Python, Rust and Go, with Kafka, Kubernetes and ClickHouse underneath.
Is he available, and where?
Open to Head of AI Engineering, CTO and AI engineering leadership roles — and fractional engagements. Based in Dubai; open to roles in the UAE, Saudi Arabia, the wider Middle East, Europe and remote.
How can a recruiter verify all this quickly?
Download the CV, read the project case studies, or ask your own AI assistant: add the public MCP server with "claude mcp add --transport http moeid https://moeidsaleem.com/api/mcp" and query his experience directly. The products are live and linked from each case study.
Building an AI engineering team?
Book a 30-minute call. Bring the role, the roadmap or the agent that won't behave in production — I'll tell you honestly where I'd start.