Salah Eldin Elnabrawy
Top Rated Plus on Upwork - top 3% of global talentI build AI systems that run real business operations. My work bridges advanced data engineering and generative AI: deterministic multi-agent workflows, citation-based knowledge engines that don't hallucinate, and high-volume automation pipelines - engineered for scale, security, and low latency, including fully on-premise deployments on constrained hardware.
Every claim on this page is verifiable through the linked profiles.
Experience
Founder & AI Systems Architect · Verel Systems
2026 - Present- ·Production AI engineering firm serving the Gulf, Europe, and the US - workflow automation agents, enterprise knowledge systems (RAG), Arabic/English voice AI, and independent AI system audits.
- ·Fixed-fee engagements with full source-code handover; on-premise deployment for data-sensitive industries.
AI Systems & Integration Architect · TheDar.AI
Feb 2024 - 2026- ·Engineered citation-based knowledge engines (LanceDB, Pinecone) for zero-hallucination answers over highly technical documentation.
- ·Architected supervisor-worker multi-agent state machines in LangGraph with human-in-the-loop approval gates for secure execution.
- ·Built event-driven automation (n8n, Make.com) and Python pipelines structuring data into strict Pydantic schemas for LLM consumption.
- ·Deployed quantized open-source LLMs for air-gapped and edge environments, maximizing throughput on 8GB-VRAM hardware.
What I work with
AI Orchestration & Agents
LangGraph stateful multi-agent systems, LangChain, PydanticAI, RAG architectures, OpenAI / Anthropic / Qwen / DeepSeek.
Data Engineering & Automation
n8n, Make.com, high-volume ETL pipelines, AsyncIO, distributed dataframes (Daft), event-driven architectures.
Backend, Infra & Model Engineering
Python (FastAPI), Node.js, Docker, GCP/AWS, GGUF quantization (4/8-bit), vector databases (Pinecone, pgvector, LanceDB), PostgreSQL.
Voice AI
Low-latency WebRTC voice pipelines, ElevenLabs, Whisper STT, Arabic + English speech systems.
Selected projects & open source
WinnowDB - local vector database in Rust + WebAssembly
High-speed vector database compiled to WASM, running natively in Node.js and the browser with no external servers. HNSW indexing at 20,000+ queries/second with sub-millisecond latency - built for privacy-first, offline RAG.
Deterministic multi-agent email automation
Production supervisor-worker LangGraph workflow automating CRM retrieval and email generation, with stateful checkpointing and mandatory human-in-the-loop approval before any send.
Workflow Automation Suite - n8n & Make.com orchestration
Event-driven workflows connecting CRMs, databases, and LLM endpoints; custom nodes and Python webhook services embedding AI classification, extraction, and drafting into business pipelines.
Arabic LLM optimization - Hugging Face contributor
Quantized complex Arabic LLMs (including Jais-13B) to 4-bit and 8-bit, enabling deployment on consumer-grade hardware - foundational work for affordable Arabic AI.
Education & programs
- —B.Eng. in Electronics & Communication Engineering, Zagazig University - the hardware-level foundation behind VRAM-optimized model deployment.
- —UAE Space Agency: Arab Space Pioneer Programme (with Space Foundation, US).
- —Egyptian Space Agency (EgSA): Technical Media Production Internship.
Work with me
Every Verel Systems engagement starts with a 30-minute call with me - not a sales team. A written plan and fixed fee within 3 days.
Book a call with me