PPO from scratch
Proximal Policy Optimization implemented from first principles in PyTorch: GAE, the clipped surrogate objective, value loss and entropy bonus.
Machine Learning & Applied AI Engineer | RL, agentic systems, enterprise AI
I build reinforcement learning systems from scratch, ship agentic platforms on Claude and MCP, and help enterprises move AI from pilots into production. I also build Arabic-first AI products.
Proximal Policy Optimization implemented from first principles in PyTorch: GAE, the clipped surrogate objective, value loss and entropy bonus.
PPO scaled across multiple GPUs with PyTorch DistributedDataParallel and torchrun, with synchronised gradient updates and sharded rollout collection.
A Gymnasium environment for training and evaluating LLM tool use as a reinforcement learning problem - agentic RL with structured tool-call rewards.
Rust + PyO3 primitives for the RL hot path - return computation and a replay buffer exposed to Python, removing per-step overhead from training loops.
A containerised ML evaluation service with FastAPI and Kubernetes (k3s) manifests - run policy evaluations as a scalable cluster workload.
My MSc dissertation work on PPO - careful study of the algorithm, its implementation details and the design choices that actually move performance.
A production-style Claude application: an MCP server, sub-agents and agent skills, built as a Forward Deployed Engineer reference for how to ship real Claude systems.
An MCP server exposing Arabic text tools - normalisation, transliteration and dialect handling - so any MCP-aware model can process Arabic correctly.
A 14-agent orchestration platform: mailbox-based messaging, persistent memory, MCP tools and human-in-the-loop control across a fleet of coordinating agents.
An autonomous Arabic video pipeline - scripting, generation and assembly chained into an end-to-end agentic workflow that produces finished content.
An AI agent that evaluates and scores resumes against a role, with structured rubric-based judgements rather than a single opaque score.
An Azure OpenAI natural-language-to-SQL analyst for a tier-1 GCC bank - lets analysts query enterprise data in plain language with governed, auditable SQL.
Arabic Microsoft Teams transcript analytics on Azure OpenAI - turning meeting transcripts into searchable, summarised, action-oriented intelligence.
An enterprise AI-adoption playbook: framework, templates and a metrics dashboard for taking organisations from curiosity to measurable production use.
An anonymised enterprise AI-adoption case study with reusable templates - what actually changed behaviour, written up honestly.
An open-source bilingual Arabic AI-literacy curriculum - practical material for getting Arabic-speaking teams genuinely capable with AI.
Arabic medical PII detection - identifying and protecting personally identifiable information in Arabic clinical text for safe downstream AI use.
An Arabic food-AI app - meal understanding and guidance built Arabic-first, currently in active development.
An Arabic modesty-fashion app built around a compositional Coverage Engine that reasons about garments and coverage. In active development.
An on-device computer-vision basketball coach that analyses shooting form from video and gives feedback - all running on the phone.
A Flutter IoT app connecting mobile to hardware - cross-platform control and monitoring for connected devices.
A creative copilot for video editing (TypeScript) - contributed work on a browser-based AI-assisted editing tool.
A free, open-source app locker (Swift) - lock apps behind an NFC tag or QR scan as an alternative to commercial focus tools.
App site, pages and examples for Nseet - landing page, privacy policy and support material for the product.
Save and restore Claude Code terminal sessions - window positions, tabs, SSH connections, models and conversations. One click to freeze before shutdown, one to restore.
A small set of lead-magnet pages (HTML) for the Mufeed brand - lightweight, self-contained marketing assets.
Exploratory notebooks on machine learning applied to trading signals - feature engineering and model experiments in Jupyter.