Led an end-to-end ML pipeline to forecast air cargo tonnage using Nixtla and Sklearn, including data pipeline creation, feature engineering, model development, AWS, and dashboard deployment with customized, business-specific explainability using Dash. Collaborated with clients weekly to align model updates with business needs, ensuring seamless client adoption. This solution reduced the time required for the financial team to generate forecasts from 6 weeks to just a few minutes, significantly improving efficiency and decision-making.Successfully delivered several advanced applications, moving beyond proof-of-concept work. These include:Chatbots based on Retriever-Augmented Generation (RAG): Leveraging LLMs to power intelligent, context-aware chatbots that provide real-time information and support across various domains.LLM Agents for Real-time Data Integration: I have built LLM-based agents integrated with databases, APIs, and files to retrieve fresh data and statuses, particularly in the logistics sector, streamlining operations and enhancing decision-making processes.Data Extraction from Documents: Using LLMs for automated data extraction from a variety of document types, significantly improves efficiency and accuracy.Cloud and On-premise LLM Deployments: Managing both cloud-based and on-premise LLM solutions to ensure flexibility, scalability, and security for sensitive data.Document Generation: Developing solutions for the automated generation of business-critical documents such as proposals, RFPs, BRDs, and test cases, ensuring high quality and consistency in deliverables.Technical Mentorship and Training: Delivered Generative AI training to internal teams and clients, building expertise in advanced AI tools and methodologies.

Spearheaded architectural optimization, converting the backend from monolithic to microservices to reduce I/O and enhance efficiency. Enabled parallel and distributed processing with Celery, RabbitMQ, and GPU acceleration, supporting high-speed document extraction. Redesigned deep learning pipelines for auto-training on new data and integrated LLMs to enhance field-level accuracy, improving reliability for banking and finance clients. This optimization reduced development times by 60%, from 5 days to just 2 days, and processing times by over 83-86%, from 6-7 seconds to under 1 second, significantly accelerating both deployment and performance.

End-to-End Fraud Detection Model Development: Independently led a bonus abuse fraud detection project, collaborating with domain experts to build a high-accuracy machine learning solution. Created training data and engineered features using SQL and BigQuery, aligning closely with business needs.Model Development and Deployment: Completed EDA, feature selection, and model evaluation using Sklearn, deploying the model through production pipelines with Docker and Airflow DAGs to ensure scalability and reliability. Custom Prediction Threshold Optimization: Developed a custom formula to optimize prediction thresholds, enhancing detection accuracy by aligning technical outcomes with business priorities.Strategic Data Warehouse Enhancement: Authored a case study recommending key feature additions in the data warehouse, automating feature creation and reducing manual data entry for long-term infrastructure improvement. The resultant solution achieved 95% accuracy in production environments when tested on live data, ensuring high reliability for real-world fraud detection.