I'm Ahmed, an architect and engineering lead based in Sydney. I design and run agentic systems: multi-agent workflows, spec-driven development with AI, and production-grade agentic apps on Azure and .NET. I treat agents the way I treat engineers. Give them clear specs, good tools, tight feedback loops, and an environment they can be productive in.
The fun part of this era isn't the model. It's the engineering around it. I spend my time designing agentic apps, wiring up tool-using agents, and turning "AI assistance" into something a product team can actually ship to customers. I bring 15+ years of cloud architecture and .NET engineering with me, so the agents I build live inside real systems, with auth, observability, evals and guardrails.
An agent is just a junior engineer with infinite patience and zero context. The leverage comes from how you brief it, what tools you give it, and how you close the loop. Specs over prompts, evals over vibes, observability over hope.
That mindset, combined with solid cloud foundations, is how I help teams move from one-off AI demos to agentic products that hold up in production.
Four overlapping practices. Pick the one that maps to where you are today.
Designing the bones of agentic products: agent topologies, tool surfaces, memory, retrieval, human-in-the-loop checkpoints, and the boring-but-vital glue (auth, queues, idempotency, retries) that makes them safe to ship.
Treating agents as collaborators: writing the specs, defining the roles, choosing models per task, setting up sub-agents and review loops, and building the workflows that let humans stay in charge of intent while agents do the heavy lifting.
Standing up the feedback loop most AI projects skip: structured evals, tracing of agent runs and tool calls, regression gates in CI, and guardrails so quality doesn't quietly slide as prompts and models change.
Turning fuzzy ideas into specs that both humans and coding agents can act on. I've built and open-sourced templates for it (see spec-pilot), and use the same patterns with Claude Code, Copilot and similar tools on client work.
From hands-on engineering to architecture leadership. The foundations behind how I build with agents today.
Designing cloud and agentic application architectures for enterprise clients across Australia. I lead architecture for Azure-based platforms, embed AI agents and copilots into existing systems, and partner with engineering teams to land production-grade, well-governed implementations.
Led software and cloud engagements end-to-end, from discovery through to production rollouts. Worked across regulated industries, balancing delivery pressure with the architecture rigour the platforms needed to last.
Hands-on full-stack and cloud engineering for a wide range of clients. Owned design and delivery of features across the stack, mentored junior engineers, and helped teams adopt cloud-first patterns.
Led a team of engineers through the full software lifecycle. Set technical direction, raised the quality bar, and architected solutions for international clients while coaching the team along the way.
Public projects I maintain on github.com/ahmedyoussef-au. Mostly tools that sharpen how I work with AI agents and ship agentic apps.
My take on managing coding agents like a team. A set of Copilot- and Claude-ready prompt templates that walk an AI through every stage of Spec-Driven Development (specify, plan, implement, review, refine) so agents produce work you'd actually merge instead of one-shot guesses.
A VS Code extension that converts Markdown files to DOCX, HTML and PDF using Pandoc, directly from the editor. Handy when an agent's output lives in Markdown but stakeholders want a polished doc on their desk.
End-to-end solution for running self-hosted CI agents/runners for Azure DevOps and GitHub Actions inside Docker containers. Ships with Bicep templates for Azure deployment. It's the same kind of foundation I use when AI coding agents need a private, controlled environment to execute in.
A minimal starter for spinning up AI and agentic prototypes with Streamlit in Python. My go-to for proving an idea quickly. Wire it up to an LLM or agent, stand up a UI in an afternoon, and iterate before committing to a full product build.