Senior Applied Scientist at Arabic.ai (Tarjama&). I build production LLM systems end to end — multi-agent workflows, RAG over dense enterprise documents, and the evaluation and observability layers that make their output auditable.
Most of my work has been for public-sector, financial services and Big Four clients: taking a vague business problem, scoping it with the people who own it, and shipping the full-stack system they actually use.
Focus right now: agent runtime safety · LLM evaluation · MLOps
Studying: MSc Artificial Intelligence, University of Leeds (2026–2028)
computer-use-agent-manager — Session
manager for concurrent Claude computer-use agents. Isolated Docker container per session with
dedicated noVNC ports, PostgreSQL advisory locks for atomic port allocation, bidirectional
WebSocket protocol for live agent events, and lifecycle management (health checks, zombie
container cleanup, graceful shutdown).
FastAPI Next.js PostgreSQL Docker WebSockets
LLMs_from_Scratch — Transformer and LLM
internals implemented from the ground up, working through the architecture one component at a time.
PyTorch Jupyter
Edge_Waste_Classifier — My BSc graduation
project: real-time waste classification running on a Raspberry Pi 4B, trained and optimised to run
inference on the device rather than in the cloud.
PyTorch Computer Vision Edge
- LLM & agents — LangGraph, LangChain, LlamaIndex, DSPy, RAG, tool calling, Langfuse, evaluation harnesses
- Backend — Python, FastAPI, PostgreSQL, Redis, Celery, Next.js, WebSockets
- Infra — Docker, AWS (EC2, S3, RDS), Azure (AI Search, AI Vision, Functions), Terraform, Airflow