420 paths, one shared factor in gold. Scroll, and the noise collapses onto it.
AI Platform & Infrastructure Lead · Abu Dhabi
I build AI platforms the way mission-critical systems should be run. Six years leading infrastructure at the Presidential Court. MSc in Computer Science (AI) at Khalifa University.
For six years I've kept mission-critical systems running. Now I build the platforms AI runs on: self‑hosted, observable and governed. Same discipline. New workload.
MSc work at Khalifa University: machine learning applied to pricing, trading and valuation. Results are reported as measured, including where a method gave something up.
At each of 49 exercise dates, a small network reads the price of every asset in the basket and estimates what holding the option is worth. If exercising now pays more, the model exercises.
d inputs → 64 → 64 → 32 → 1 · ReLU · dropout 0.1 · Adam · one network per exercise date
We replaced the regression step of Longstaff–Schwartz Monte Carlo with a PyTorch neural network, and benchmarked it against the classical polynomial method on baskets of 1 to 100 correlated assets with 50,000 simulated paths.
39 s instead of 1.9 hours, with prices within $0.50 up to 15 assets and within $2 at 100.
| Assets | Polynomial (s) | Deep (s) |
|---|
A trading environment with costs, turnover penalties and position limits. We trained PPO, DQN and a custom Double DQN, with a strict time-based split so the agents never saw the test period.
PPO earned less, but with far less risk: a higher Sharpe ratio and a third of the drawdown.
| Strategy | Return | Max drawdown | Sharpe |
|---|---|---|---|
| PPO agent | 11.8% | −4.0% | 1.61 |
| Buy and hold | 20.1% | −12.9% | 0.90 |
| Momentum | 11.1% | −8.4% | 0.80 |
A machine-learning pipeline on 1.19 million Connecticut property sales, cleaned and audited down to 558,000 arm's-length residential sales, with any field derived from the sale price removed so the model can't cheat.
The model never learns from the future. Linear, regularised and tree-based models are being benchmarked next.
Systems I've built in my own lab: private models, agents that run offline, and the plumbing that keeps them reliable.
A private, fully on-premises LLM service. Open-weight models run on two inference hosts behind a LiteLLM gateway that load-balances, falls back when a host is down, and exposes one OpenAI-compatible API.
The lab underneath (Proxmox, kubeadm Kubernetes) is built and rebuilt end-to-end with Ansible.
A multi-agent system (orchestrator, planner, task manager, code and shell agents) running entirely on local models, with a sandboxed tool layer, approvals and undo.
Local models lacked native function calling, so I implemented text-based ReAct tool calling: the agent writes a structured request, the system parses and runs it, and the user only sees the result.
A modular pipeline from market and news ingestion through momentum and volatility factors, FinBERT news sentiment, volatility-targeted allocation and backtesting.
An agent layer (screener, risk, and a bull-versus-bear debate) places orders through Interactive Brokers paper trading, with a monitoring dashboard.
From an ADIA scholarship to leading the platform behind mission-critical services, and now into AI.
MSc Computer Science, Artificial Intelligence concentration (part-time)
Systems & Virtualization Lead (Team Lead)
BSc Computer Science, Minor in Mathematics · ADIA Scholarship
Software Engineer Intern · USA
Developer Intern
Happy to talk about AI platforms, sovereign infrastructure and governed AI.