Building HealthQCopilot — an agentic AI healthcare platform on Azure with Semantic Kernel, Phi-4-Medical, and FHIR. Eighteen years scaling enterprise SaaS across healthcare, fleet, trade finance, and government, and serving on the Architecture Committee at Holman.
The career path that lands at AI-first architecture. Eighteen years of enterprise platform delivery and architecture ownership, now compounding into Principal-level direction-setting for agentic AI, micro-frontend, and cloud-native systems.
Senior member of the Architecture Committee for the Remarketing Solutions, Supply Chain, and Fleet & Maintenance Engineering organization. Co-own multi-year platform strategy for the systems that power 300+ enterprise fleet clients and roughly one million transactions per day on Azure Container Apps, Service Bus, Cosmos DB, and Azure APIM.
Architected distributed integration patterns supporting $5M+ in vehicle sales. Pioneered Dapr adoption — 40% inter-service latency reduction. Built serverless ETL pipelines processing 10,000+ events per second with WAF-fronted, audit-logged API governance.
ASP.NET Core microservices on Azure Container Apps with KEDA autoscaling. Micro-frontend React platforms via Webpack 5 Module Federation. Lifted automated test coverage from 45% to 85% as an org-wide standard.
RESTful API platform for vehicle remarketing. Centralized observability with ELK and OpenTelemetry reduced mean-time-to-detect by 60%.
Solution architecture and .NET / Azure delivery for boutique clients across Dhaka and the EU corridor. Identity integration with OAuth2 / OIDC; multi-tenant SaaS patterns.
Trade finance microservices for DHL, ING Bank, and Aronova. .NET Core, CQRS, RabbitMQ, IBM Hyperledger Fabric. Multi-tenant identity and API gateway patterns for 100+ enterprise clients.
The HealthQCopilot agentic AI platform, a Module Federation micro-frontend shell, and a generic GraphQL gateway reference pattern. Hover the nodes — each one has a note about what runs inside.
A HIPAA-grade agentic platform built on .NET 10 and Semantic Kernel. A patient encounter — voice plus structured intake — is routed through five specialised agents around an orchestrator: triage, clinical coding, prior authorization, care-gap detection, and a hallucination guard that verifies every output against the patient’s FHIR record before it reaches the EHR.
The pattern for splitting a monolithic React app into independently deployable remotes. A thin shell host owns routing, auth, design system, and telemetry; each remote ships its own Vite/Webpack bundle, exposed at runtime via Module Federation. Independent teams, independent deploys, one cohesive product.
The pattern I’ve used to collapse fragmented REST surfaces into a single governed schema on .NET + HotChocolate behind Azure APIM. OAuth2 with field-level authorization through GraphQL directives, edge throttling at APIM, and event-driven services behind the gateway.
Eighteen years of compounding depth across cloud, distributed systems, and now agentic AI — the toolset that ships the platforms above.
Two long-running ventures alongside the principal role — both released under permissive open-source licenses so the work compounds beyond any single employer or geography.
Healthcare is too important to lock behind a single vendor. HealthQCopilot is being released as open source so any hospital, clinic, or public-health agency — not just well-funded systems — can deploy HIPAA-grade agentic AI on their own infrastructure.
Built on .NET 10 and Semantic Kernel: an ambient scribe captures the encounter, five specialised agents reason over it (triage, clinical coding, prior-auth, care-gap, and a hallucination guard), and every output is verified against the patient’s FHIR record before it reaches the EHR. Nine microservices behind a YARP gateway and GraphQL BFF.
Independent research building Bangla-language LLMs for social-impact applications — including migrant-worker protection and education access. Closing the gap for one of the world’s most-spoken but most-underserved languages in foundation models. Data, evals, and model weights are being released openly so other Bangla-speaking communities can build on top.
Pieces I publish from the architecture work. Each one starts with a failure, then the fix.
Why a hallucination guard against FHIR became the most important agent in the system, what we got wrong about prior-auth, and the orchestration patterns that survived contact with real clinicians.
Read →Why the registry and design-system packages matter more than the remotes themselves, the runtime errors that nearly killed the rollout, and the team-topology pattern that made independent deploys actually safe.
Read →The pattern for consolidating an enterprise platform onto HotChocolate behind Azure APIM — including the directive design that finally made field-level authorization survive a security audit.
Read →I’m currently exploring remote and hybrid Principal / Staff Architect roles where cloud-native platform thinking, enterprise AI, and pragmatic engineering leadership compound. PE-backed SaaS, fleet / logistics, fintech, and AI-platform companies feel most aligned.