New Delhi, Delhi, India
Lead AI/ML Engineer and Data Scientist with over 6 years of experience specializing in developing AI solutions, including Personalized Recommendations, User profiling, Large Language Models (LLMs), Multimodal RAG systems, and scalable AI platforms. Proven expertise in building robust, production-ready AI capabilities from scratch, facilitating seamless AI adoption by software development teams. Skilled in end-to-end Machine Learning project lifecycles—from research, prototyping, and model development to cloud-based deployments and optimization. Passionate about leveraging cutting-edge AI to solve real-world problems and accelerate intelligent automation across domains.
Enterprise AI Platform and Agentic Automation for C2C workflows: Architected and deployed a multi-agent agentic AI system using CrewAI,Langchain, and Llama-Index to automate manual invoice processing and compliance workflows across enterprise customers. -Technical approach: Designed stateful agent orchestration with human-in-the-loop (HITL) checkpoints, enabling operators to adjust agent autonomy dynamically based on task risk levels. Implemented memory management for sequential task execution without hallucination drift and toxicity. -Tools & stack: GPT-4, Llama-Index, LangChain, CrewAI for orchestration, PySpark for batch processing, DeepEval for evaluation. -Business impact: 36% reduction in manual processing resources (~$420K annual savings), 45% faster invoice processing time, freed ~12 FTE hours/week for higher-value work. -Key technical challenge solved: Agents initially struggled with complex multi-step document extraction, hallucinating loops, and task failures. Implemented a custom retrieval-augmented generation (RAG) layer that indexed structured data from source systems, reducing extraction errors from 15% to <2%. Created agent structure with CrewAI with HITL checks to solve logical failures and used DeepEval metrics to evaluate LLM responses. Multi-Modal Knowledge Graph : -Designed and deployed a RAG system enabling employees to search thousands of complex PDF manuals (manufacturing, retail, healthcare domains) using natural language queries. -Technical architecture: Built a document embedding pipeline with multiple modalities (text, tables extracted via OCR, images processed with vision encoders). Used hierarchical chunking to preserve context while keeping vector database queries efficient (<5s for 10K+ documents). -Business impact: Reduced average document lookup time from 17 minutes to <5 minutes per query, improved first-contact resolution rate by 32% (~$100K annual support cost savings), and accelerated onboarding by ~60 hours per new hire.
Worked in the core Data Science Practice Team at Inshorts.We did a lot of cool experiments like: 1. A User Profiling and News Recommendation Engine :Built a personalized dynamic news recommendation engine algorithm for close to 100 million users modeled by a graph based inferential approach and user profiling through a cluster based distribution analysis over their individual timespents spread across active articles at any given moment. 2. A News Notification Engine : Developed an efficient notification engine to deliver daily news updates, leveraging a powerful news classification algorithm. The implementation led to a remarkable 6.8% increase in average DAU, driving significant inorganic user engagement. 3. A News Categorization, Summarization and Tagging Engine: Worked on implementing sharp generative AI models to automate news summarization and tagging. The solution would generate concise 60-word conceptual summaries and keyword tags, eliminating the need for manual intervention. These automated summaries will facilitate seamless article search and collaborative content aggregation.
Worked in the Data Science Practice team in the Centre of Excellence (COE) group of the organization. Responsible for researching, developing and implementing solutions to real world business problems across a wide range of business domains with an emphasis on the areas of Machine Learning, Decision Optimization, Predictive Modeling, Forecasting, Metaheuristics, Deep Learning 1 ) Personalized Recommendation Engine : Built a product recommendation engine for boosting store sales through customer interaction and targeting customer retention through personalized offers. Increasing customer cart values using recommendations based on collaborative filtering, content based filtering, graph theory, guided selling and customer segmentation. 2) Last Mile Delivery Optimization: Optimized last mile delivery with multiple hubs delivering goods for multiple customers in a city in their preferred time window. The solution helped minimize the operational costs, maximize resource utilization and serve customer demand on time. 3) Inventory Management : Built reinforcement learning agent solution to determine the optimal level of inventory for thousands of products. The solution prevents understocking and overstocking leading to better utilization of the capital invested and improving order-fulfillment by delivering within the committed lead times. 4) Real Time Flexible Scheduling : Created a meta heuristic solution using evolutionary algorithms for optimal scheduling of large scale manufacturing orders. The solution reacts to dynamic changes in the manufacturing environment such as inflow of orders, machine breakdown, resource unavailability, production delays etc to create an optimal job schedule. It minimizes the makespan, tardiness(total delay) and maximizes the resource utilization.
Developed explainable machine learning models to detect potential loan defaulters for a leading bank in Singapore.. The models rendered actionable insights to evaluate credit risk strategies and mitigate bad loans.