Data Scientist
Linde Engineering India · Aug 2023 – Present
- Architected a global ML-powered software license optimization platform supporting 2,000+ engineers across multiple business units; developed XGBoost forecasting models in Azure Machine Learning (~13% MAPE) using historical utilization patterns, seasonal demand patterns, and team-level consumption behavior, while OR-Tools constraint optimization dynamically allocated licenses, reducing enterprise software licensing costs by 25% and improving license utilization by 22%.
- Owned production ML infrastructure supporting enterprise license allocation decisions; engineered Azure Databricks (PySpark) and Airflow pipelines processing 500K+ daily usage records, deployed Azure-based Docker inference services, and implemented MLflow model registry, automated drift monitoring, retraining pipelines, and SHAP explainability dashboards to enable reliable, scalable, and transparent ML operations.
- Built a predictive supply chain intelligence platform supporting procurement and operations teams across 10+ regions; processed 8M+ procurement and operational records and developed XGBoost predictive models integrating consumption patterns, supplier behavior, demand signals, and lead-time features, achieving 85% model accuracy through cross-validation while reducing material waste by 27% and over-ordering by 32%.
- Developed a procurement risk intelligence platform across 25+ supply categories; built automated risk scoring models using supplier reliability, demand volatility, inventory constraints, and purchasing behavior, reducing manual procurement analysis by 40% and enabling proactive supplier risk identification before sourcing decisions.
- Designed a secure enterprise knowledge retrieval platform indexing 300+ internal documents; implemented production RAG pipelines using Azure OpenAI, embedding-based retrieval, Pinecone vector indexing, hybrid search, and document-level access controls, enabling employees to query unstructured knowledge sources through natural language and reducing manual document search efforts.
- Optimized enterprise RAG performance through retrieval evaluation, embedding selection, document chunking strategies, and LLM response tuning; implemented evaluation workflows measuring precision@K, recall@K, and response faithfulness, improving retrieval precision and answer accuracy by 30% across knowledge management workflows.
- Engineered an autonomous Due Diligence Agent orchestrating multi-source web research and synthesizing financial risk, contract, and compliance assessments into structured PDF reports, replacing a manual workflow requiring 4+ analyst-hours per entity and accelerating investment review processes.
- Deployed secure RAG pipelines indexing 100+ confidential financial and business documents, enabling finance teams to perform natural language analysis over unstructured PDFs while maintaining document-level access controls, audit logging, and compliance requirements.
- Fine-tuned Llama models with QLoRA on domain-specific financial terminology, improving extraction accuracy by 15%; deployed optimized inference through FastAPI and production pipelines, reducing latency for financial analysis and reporting workflows.
- Designed a Financial Comparison Agent that benchmarks project and business performance metrics and automatically generates management reports, reducing manual comparison efforts by 60% for financial reviews and cross-functional stakeholder discussions.


