ziad.makkawi
Portrait of Ziad Makkawi

L0 input
Nottingham, UK

Ziad Makkawi

Software engineer working down the stack, from backend systems toward ML inference, and relearning C++ to get there.

toy serving loop · continuous batchingqueue 0 · batch 0/6 · 0 tok/s

L1hiddenAbout

I'm a software engineer in Nottingham. I'm a Technical Consultant at XMA, where I configure HaloPSA, automate workflows and deploy ITSM platforms for clients. In my spare time I'm the founding engineer at LeedAB, which builds AI-powered trade operations for physical commodity exporters.

Lately I've been heading further down the stack, toward ML inference, the part that comes after a model is trained, where speed and size matter. That's why I'm relearning C++ properly. The lab below is the first result.

I graduated from Nottingham Trent University with a first-class degree in Software Engineering. My final-year project was a CNN that classifies brain tumours from MRI scans.

L2hiddenExperience

  1. Technical Consultant · XMA

    Jan 2026 – Present

    Nottingham, UK · Hybrid

    • Deliver technical consultancy and implementation services to clients across IT service management and professional services, specialising in HaloPSA configuration, workflow automation and ITSM platform deployment.
    • Work closely with clients to understand their challenges, scope solutions and deliver outcomes across onboarding, system configuration and ongoing support.
    • Contribute to the buildout of XMA’s Professional Services practice from the ground up.
  2. Founding Engineer · LeedAB

    Jun 2026 – Present

    Remote · Part-time

    • Founding engineer of an AI-native trade operations platform for physical commodity exporters, bringing the documents, messages and shipment data traders work across into one system.
    • Design and build LLM agents that verify trade documents against contract and letter of credit terms and flag discrepancies before they cause rework or delays.
    • Build the integration layer connecting messaging, email and trade tools so agents can act across a trader’s existing workflow, plus shipment tracking across trade corridors.
    • Work with commodity exporters and traders to shape the product around how trades actually run.
  3. Software Engineer Intern · NHS

    Aug 2024 – Sep 2024

    Nottingham, UK

    • Boosted simulation tool accuracy by validating patient flow models in Python, improving resource allocation by 20% across scenarios.
    • Wrote tutorials and guidelines on GDPR-compliant practices for public version control such as GitHub, adopted by the team.
    • Presented workflow automation and testing outcomes at the NHS-R community webinar.
  4. Software Engineer Intern · TheirCircle

    Aug 2023 – Mar 2024

    Toronto, Canada · Remote

    • Implemented responsive UI components and dynamic feature modules, increasing user engagement by 35%.
    • Optimised backend API integration, improving data retrieval performance by 40% on iOS and Android.
    • Built and deployed a CI/CD pipeline that automated testing and deployment, cutting release times by 25%.

Education

BSc (Hons) Software Engineering, First Class · Nottingham Trent University

2020 – 2023

L3hiddenProjects

L4hiddenLab

int8 cuts the model from 199 KB to 50 KB for a 0.01-point accuracy drop, and the C++ engine that runs it compiles to 1.5 KB of WebAssembly.

A small model running in your browser. Draw a digit and it classifies it on every stroke. The weights are quantised to int8, and the inference engine is written in C++ for WebAssembly, with a TypeScript twin as a fallback.

draw a digit 0–9

prediction

?

  1. 00%
  2. 10%
  3. 20%
  4. 30%
  5. 40%
  6. 50%
  7. 60%
  8. 70%
  9. 80%
  10. 90%
engine loading… · 784→64→10 MLP, int8 weights · latency – (mean of 200 runs, measured in your browser)

size 199 KB · accuracy 97.83%

64 real weights from the model above. Lower precision makes the model smaller and faster, until it doesn't work any more: int4 keeps almost all the accuracy, int2 falls apart. Accuracy is on the 10,000-image MNIST test set.

L5hiddenNow

LeedAB, on the side
In my spare time: AI document checking, a buyer portal and a trade assistant for physical commodity exporters.
C++, again
Relearning modern C++ from the ground up, with an eye on performance. The lab’s inference engine is the first result.
ML inference
Learning what happens after training: quantisation, serving, and making models run fast and small.

Completed Boot.dev’s backend curriculum (Aug 2025 – Jan 2026): Python, Go, SQL, HTTP servers, Docker, system design and AI agents.

toolbox: Python, C++, TypeScript, Java, C#, Go, PostgreSQL, TensorFlow, WebAssembly, React, Next.js, Docker, Git, LLM agents, MCP, HaloPSA, Flutter, .NET Core

L6outputContact

Open to conversations about backend, ML systems and inference work.