user1824201

Developed production multi-agent AI systems for claims leakage auditing and claim summarization. Built agent workflows using LangChain Deep Agents on LangGraph, reducing manual review time by 90% for leakage auditing and improving finding reliability via domain-aware actor–critic, source attribution, and human-in-the-loop verification across 100 ground-truth claims. Integrated 18 live claim/policy Model Context Protocol (MCP) tools with caching, retries, and telemetry, and engineered resilient agent harnesses with Pydantic-validated structured outputs, retry/timeout controls, and partial-result recovery. Implemented policy context management for 200+ page documents using tool-result offloading and actor–critic preclassification while reducing latency. Created LLM evaluation and observability using OpenTelemetry and Azure Application Insights for trace-level monitoring of agent/tool latency, failures, tokens, and cost, and deployed a Streamlit app via on-prem Kubernetes and Azure De

Delivered end-to-end ML and GenAI solutions in healthcare focused on speech analysis and clinical text generation. Built ML pipelines achieving 80% F1 for slurred speech detection and deep learning models reaching 85% F1 using PyTorch, with MLflow tracking and FastAPI/Docker deployment on GCP. Integrated explainability with SHAP for transparent clinical decision support. Fine-tuned BioMistral-7B using LoRA with 4-bit quantization in PyTorch/Hugging Face, tracking 50+ runs in Weights & Biases. Developed a Retrieval-Augmented Generation pipeline to help paramedics access clinical guidelines.

Improved image classification performance by developing teacher–student optimization methods for knowledge distillation. Co-authored a NeurIPS Workshop 2025 paper on logit-based losses limiting feature knowledge distillation. Achieved faster convergence (about 80% faster on CIFAR/Tiny ImageNet) using a knowledge quality (KQ) metric to automatically select optimal teacher layers, implemented in PyTorch. Improved CNN/ViT accuracy by about 15% via a logit-free KD framework.