Ehsanul Haque

Azure Platform Insight is a production multi-agent orchestration pipeline built on Semantic Kernel and Azure AI Foundry that analyzes IcM incidents to determine whether Azure platform failures caused customer-facing issues.

As the principal AI engineer on the ACES/ACE team, I own the technical direction and system architecture. My responsibilities included designing the multi-agent orchestration architecture and the RAG pipelines that grounded agents in Azure incident, wiki, and TSG knowledge, as well as the long-term memory subsystem and its guardrails that keep agent reasoning consistent and safe. I architected the MCP server layer and its authentication design (MISE, ARM RBAC) to securely expose diagnostic tools.

I also led the evaluation and observability strategy, building evaluation scaffolding and feedback pipelines that continuously measure accuracy, detect regressions, and surface failure patterns, such as evaluator cascades and synthesizer misrepresentations. I also authored technical design documents and tracked framework convergence between MAF and Semantic Kernel to keep the production system current.

IBM iConnect Access is a zero-download, web-based enterprise medical imaging platform that enables healthcare providers to aggregate, exchange, view, and collaborate on diagnostic imaging studies across the enterprise and extended care teams from any device. As the senior ML engineer, I designed and deployed the AI systems, bringing generative and agentic capabilities into this regulated healthcare environment. My responsibilities included building a real-time, GPU-accelerated agentic AI pipeline on Kubernetes, orchestrating LangChain and LangGraph LLM agents for autonomous task coordination and embedding generation, and managing vector databases, with streaming via Kafka and Flink, model serving on Triton, and observability through Prometheus and Grafana.

I architected a RAG system that combines LLMs with advanced retrieval using vector embeddings to improve accuracy in question answering and summarization. I fine-tuned transformer-based LLMs (BERT, GPT, Llama) using QLoRA, mixed precision, and gradient checkpointing, scaling distributed training with DeepSpeed and model/data parallelism. I also built automated training and evaluation pipelines and deployed models to production via CI/CD and containerization.