supre3th

Built a Hierarchical Multi-Agent system with LangGraph, featuring a Supervisor orchestrating worker agents to emulate Data Engineering workflows and enable agent interoperability. Developed a Self-Reflecting subgraph with validator, refiner, and executor GenAI agents for automated DML/DDL processing, leveraging a Vector DB ( RAG for Few-shot) for error correction and continuous learning from executor and Human Feedback. Improved query reliability by 40% and reduced user setup time by 60% by minimizing manual query fixes. Designed the framework for adaptability, enabling rapid extension to new personas by simply defining tools and agent descriptions. Delivered the first offshore-driven GenAI solution at LSHC for intent classification, resolving an estimated 11 million missed opportunities annually and achieving 80% Precision. Led a team of 3 in exploration of Zero-Shot, domain Agnostic GenAI algorithm for detecting LLM Generated content in clinical notes and bills, achieving 96% Precision and 97% Accuracy to prevent fraud on GCP Vertex AI. Developed a proprietary OR Optimization Algorithm leveraging PuLP, Integer Programming, and Machine Learning models (e.g., logistic regression for time estimation/wastage, Item-Based Collaborative Filtering for room recommendations) to optimize schedules, secured $600k+ in investment, and driving adoption by 1 client and installation deal with another after demos to 8 clients. Architected the Health Assist GenAI suite, integrating Retrieval Augmented Generation ( RAG ) and text To SQL capabilities built using Langchain to empower CSR agents with rapid access to patient benefits and data, streamline query resolution, and enable AI-guided conversations for enhanced service efficiency and patient experience.

Optimized patient email Intent Classification Accuracy from 48% to 85%, reducing Inference time from 30s to <1s via Knowledge Distillation and Quantization, and cutting Training time from 4–5 days to under 7 hours through Parallel GPU training, significantly enhancing operational efficiency. Enhancement, Refactoring and Modularization of Feature store consumption Python SDK, currently being used by the client to store more than 200+ features. Development of MLOps capabilities on Azure Kubernetes using Kubeflow and a customizable, declarative ML/DL lifecycle pipeline enabling rapid prototyping reducing efforts requiring 1 month to a week. Implementation of an AI/ML model testing and AB testing framework aligned with industry best practices; ; developed a custom Python SDK for Azure Storage to support framework functionality. A projected incremental value of $6M–$14M. Automated NLP and NER-based keyword extraction by analyzing policy context and generating top keywords; enhanced Universal Dictionary to extract concepts, map terminologies, and support partial text match, semantic Similarity, and full text match for advanced knowledge retrieval. Data X-Ray Accelerator to validate document structure and content against reference templates by checking section order, subsections, and content; leveraged NER to mask document-specific keywords and used BLEU score and semantic similarity metrics to assess template adherence.

Utilized one-shot instance segmentation and fine-tuned Mask-RCNN to accurately detect water bodies, vehicles, and key features in aerial mining site imagery, significantly improving detection capabilities and analysis efficiency. Developed a framework for assessing physical rehabilitation exercises using OpenPose, Dynamic Time Warping, and Affine Transformation to automate quantification of patient performance in completing prescribed exercises. Utilized various GAN architectures for text to Image synthesis, generating realistic images from textual descriptions. Automated vehicle insurance processing by detecting damaged vehicle parts and suggesting settlements based on severity, fine-tuning YOLOv3 for object detection and ResNet50 for severity classification. Developed an XAI platform to explaining decision-making processes, utilizing Layer Wise Relevance Propagation (LRP) to interpret predictions of neural machine translation models.