Associate Technical Architect
Jun 2025 — Present- Architected and built a private AI agent platform, now in production in the client's own Azure tenant on hub-and-spoke topology, giving business units self-service deployment of agents and the web applications consuming them - private networking, PHI handling, monitoring and audit are properties of the platform rather than problems each team solves again.
- Built regulatory retention into the platform itself rather than leaving it to each application - logs held 90 days, other categories beyond 360 days, and records required for seven years - so a team shipping onto the platform inherits compliance instead of implementing it.
- Designed the access and network architecture end to end: every application deployed into a private VNet, all web traffic entering through an Application Gateway, an Azure Firewall on egress, and a Bastion service as the only route to build hosts, so engineers deploy through pipelines without the estate ever holding a public endpoint.
- Resolved a conflict between Australian data-residency law and audio models available only in East US by combining multi-region model routing, automated screening of data leaving Australia and mandatory security-reviewer approval, narrowing cross-border traffic to an audited exception path instead of dropping the capability.
- Built an evaluation framework for agents and prompts and applied it across the AI projects on the platform, establishing measurable quality baselines before release, and integrated Microsoft Fabric so agents query governed enterprise data at source rather than through duplicated copies.
- Configured content moderation, prompt-injection defence and output controls on Azure AI Foundry, and applied systematic prompt iteration across the platform's agents.
- Took over a clinical AI application delivered late by another team, worked it for bug fixing before concluding that unclear naming, uncaptured business scenarios and unfit scaling warranted rebuilding - then redesigned it as a Next.js application consuming Microsoft Fabric data, with Python agents on the Azure Agent Framework and Cosmos DB for application state.
- Directed Claude Code and its design canvas to build a clinical note-taking feature end to end for a policy-grounded care-plan assistant - spec written with test cases first, agent generating the build plan, with the plan and edge cases corrected by hand - for a second Australian aged-care client.
- Designed a three-agent pipeline (language detection, speech transcription, translation) letting multilingual aged-care staff record shift notes in their own language, with a summary and action list derived automatically for the next carer - live in production, and the multilingual note core was later adapted for a second Australian healthcare client.
- Gave the agents a governed tool surface against a real system of record: a C# console MCP server on the .NET MCP SDK wrapping the aged-care system of record, reading across a handful of endpoints and writing back exactly one record type, progress notes, by deliberate design rather than broad access.
Azure AI FoundrySemantic KernelAzure Agent FrameworkMCPRAGAzure AI SearchMicrosoft FabricLangGraphNext.jsPythonC#Cosmos DBBicep