Kabya Basu

– Spearheaded the development of a health and wellness application that analyzes user behavior data to deliver personalized, Al-driven recommendations. Designed and implemented all machine learning components, agentic Al systems, and the complete frontend, integrating RESTful APIs for seamless data exchange between frontend and ML backend. – Built a production-grade, end-to-end video compliance auditing platform for government healthcare agencies using Azure, LangGraph, RAG, and GPT-4 to automatically analyze healthcare-related YouTube videos and generate real-time, structured compliance intelligence reports. – Built a full end-to-end Retriever-Augmented Generation (RAG) pipeline to enhance the personalization and contextual relevance of Al responses, incorporating unit and integration tests to maintain 95% code coverage and ensure robust system reliability. – Designed and deployed an agentic Al system to automate user demand fulfillment in the application's beta release, utilizing GitHub Actions for automated CI/CD workflows to enable frequent, high-quality releases. – Fine-tuned a large language model using Gradient-based Representation Optimization (GRPO) and advanced prompt engineering techniques (including DSPy for dynamic prompt optimization) to enhance its reasoning capabilities, improving EHR-to-clinical trial matching accuracy by 70% over the base model. Achieved efficient distributed training by leveraging multiple GPUs via data parallelism and model sharding techniques. – Optimized large-scale Al deployments on Microsoft Azure, integrating monitoring and observability frameworks to ensure 99.9% system uptime and achieving a 20% reduction in operational costs, while conducting end-to-end testing to validate cross-system integrations.

– Led a global team of four MLOps engineers to deliver NLP, computer vision, and LLM projects, focusing on post-processing and evaluation frameworks to optimize business KPIs, including the design of RESTful APIs adhering to best practices for model inference and data retrieval. – Designed an end-to-end MLOps platform using Azure DevOps, Spark, Airflow, and GitHub Actions to orchestrate ML pipelines, incorporating CI/CD, model validation, observability, and automated unit/integration/end-to-end testing to achieve 90%+ test coverage, ensuring high reliability and scalability. – Architected and fine-tuned a T5-based model to enhance the quality of clinician address data, resulting in a 20% improvement in data accuracy and bolstering downstream data mapping workflows in healthcare applications. – Designed and operated GPU-enabled AKS clusters with separate node pools for training and inference workloads, implementing HPA and cluster autoscaler for dynamic scaling and cost optimization. – Utilized Terraform to manage and provision infrastructure as code for Azure-based ML environments, enabling automated deployment of resources and improving infrastructure consistency, while applying API design principles to create secure, scalable endpoints for model serving. – Refactored the monolithic codebase written by data scientists into a modular, well-documented, and git version-controlled structure, significantly improving maintainability, scalability, and team collaboration while optimizing production inference costs and incorporating comprehensive test suites. – Conducted technical workshops focused on improving code quality, modularity, and maintainability across the Al engineering team, including sessions on prompt engineering advancements like DSPy to keep the team aligned with the latest research in the field.

– Led a team of thirteen consultants for Walmart in designing, building, and deploying an end-to-end machine learning solution with MLOps maturity of level 3 to detect diseases from retail purchase behavior. The model achieved 93% accuracy, 80% precision, and 86% recall, with projected annual sales growth of 13%. – Delivered an analytical solution for Microsoft's Business Excellence - Ops team by understanding their data and expectations. This led to a 17% improvement in CSAT scores. – Managed the team's delivery activity through agile sprint planning, daily scrums, and being an enabler across a multidisciplinary team. – Collaborated closely with product, platform, and business teams to align Al solutions with organizational goals, ensuring the successful deployment of generative Al models across diverse industries.

– Created "algocontainer," a Python package for rapid prototyping of projects for various steps of the data science lifecycle. This package improved average project turnaround time by 23% when supervised or unsupervised machine learning algorithms were used. – Delivered pricing optimization solutions at SKU and store levels for a Fortune 1 retail store chain, using clustering, PCA, and time-series algorithms. Achieved the required accuracy in the deployed model according to the client's acceptance criteria. – Created a solution using a GAN model that replicates text patterns from successful product listings. Defined an alternative loss function as a new class of functions that forces the model's generated output to include a set of user-defined keywords. This allows the generative adversarial network to recommend rewording the phrasing of a listing description to increase the likelihood of product sales.

– Led a team of 15 scientists to develop predictive models integrating multi-omics data (genomics, transcriptomics) for drug response prediction in cancer patients, leveraging deep neural networks and preprocessing single-cell RNA-seq data. – Conceptualized the CPSA tool, an in silico platform superimposing single-cell genomics on virtual cell models, improving drug outcome predictions by 20% in clinical trials. – Clustered patient datasets using agglomerative clustering for the MyCare002 project, integrating single-cell RNA-seq and Foundation One trial data to support personalized medicine. – Published 26 international research papers on machine learning and system biology, contributing to the advancement of Al research. (https://bit.ly/4kVQLit ).