Skip to content
Book a Call

AI ENGINEER

ZEYAD DIAA

Preparing your experience…

AI EngineerCairo

ZeyadDiaa

AI engineer building AI agents, retrieval systems, and LLM applications, from data pipelines to production deployment

About me

Building systems thathold up in production

I design and deploy AI systems that connect models to real workflows. My work spans retrieval-augmented generation (RAG), AI agents, document understanding, and backend engineering. I currently build enterprise AI solutions for HR and ERP operations, taking systems from the first data pipeline through deployment and helping teams use them with confidence

Experience

Backend foundations,AI in production

  1. Jan 2026 — Present

    Building enterprise AI agents and LLM systems that automate HR and ERP operations, from prototype through to production, and training the teams who use them

  2. 2024 — Present

    Independent client work across AI and backend — LLM applications, APIs and services, workflow automation, and data pipelines and dashboards

  3. Sep 2025 — Jan 2026

    Applied AI across retrieval, document understanding and computer vision — building RAG pipelines, fine-tuning models, and automating the workflows around them

  4. May 2024 — Aug 2025

    Production backend engineering — designing microservices and REST APIs, and the data and search layers behind them

Selected Work

Systems I have built, end to end

01Multimodal RAG

PolySummMultimodal RAG for Scientific Paper Summarization

A multimodal RAG system that reads a scientific paper end to end — text, figures and tables — and writes a summary grounded in all three

  • Built a multimodal Retrieval-Augmented Generation (RAG) system that summarizes scientific papers by fusing text, figures, and tables, deployed as Dockerized FastAPI microservices behind a React frontend
  • Engineered the document AI ingestion pipeline: Detectron2 and LayoutParser layout detection with PaddleOCR (PP-OCRv3) text and table recognition converting PDFs to structured JSON, then semantic chunking and Sentence-Transformer embeddings indexed in a ChromaDB vector database
  • Fine-tuned and served modality-specific models: PEGASUS-X for abstractive text summarization (Hugging Face Trainer, ROUGE evaluation) and PaliGemma-3B with LoRA/PEFT adapters for figure captioning, plus a Qwen vision-language model for table-to-text generation

Built with

  • Python
  • PyTorch
  • Hugging Face Transformers
  • LangChain
  • FastAPI
  • ChromaDB
  • PaddleOCR
  • Detectron2
  • Docker

Concepts

  • LoRA/PEFT

02Multi-agent systems

Agentic AnalysisMulti-Agent AI Data Analyst

A crew of role-specialised agents that turns a raw spreadsheet into an executive-ready report, charts and all

  • Built a multi-agent AI data analysis system with CrewAI in which three role-specialized Large Language Model (LLM) agents run as a sequential crew with task-chained context, turning an uploaded CSV or Excel dataset into an executive-ready PDF report
  • Developed five custom agent tools (data loader, pandas-backed data query, chart discovery, chart generation, PDF report writer) enabling automated exploratory data analysis and programmatic chart generation
  • Implemented a conversational chat agent that answers natural-language questions over the dataset by autonomously selecting the appropriate tool per query, and deployed the app to Streamlit Cloud with model-agnostic LLM routing via OpenRouter

Built with

  • Python
  • CrewAI
  • Pandas
  • Matplotlib
  • Seaborn
  • Streamlit

Concepts

  • LLMs
  • Agentic AI
  • Prompt Engineering

03Retrieval over video

Video RAG AssistantQuestion Answering over Video

Ask a question about a video and get an answer back with the timestamp it came from

  • Built an end-to-end multimodal RAG assistant for question answering over video, orchestrating a five-stage pipeline: FFmpeg audio extraction, Whisper speech-to-text transcription with timestamps, OpenCV keyframe extraction, embedding generation, and vector indexing
  • Implemented semantic search over transcript segments using 384-dimensional all-MiniLM-L6-v2 embeddings with cosine similarity and top-k retrieval in a Qdrant vector database, grounding LLM-generated answers in timestamped source segments and keyframe thumbnails

Built with

  • Python
  • PyTorch
  • OpenAI Whisper
  • Sentence Transformers
  • Qdrant
  • OpenCV
  • FFmpeg

Concepts

  • LLMs
  • RAG

More builds

04 projects

Expertise

How I can help,and what it takes

A chatbot that answers from your own documents — contracts, manuals, research — and shows you where each answer came from

  • Answers grounded in your files, not guesswork
  • Retrieval-augmented generation (RAG)
  • Citations back to the source page
  • Chat over documents, audio and video
LangChainQdrantHugging Face TransformersPyTorchPython

Loading writing…

Contact

Start withthe problem

Have a role, project, or technical challenge in mind? Share what you are building, where you need help, and any timeline. I will reply with practical next steps

Cairo, Egypt

Would rather talk it through?

Book a call