Dharshini A.

Project: 𝙎𝙚𝙢𝙞-𝙎𝙩𝙧𝙪𝙘𝙩𝙪𝙧𝙚𝙙 𝘿𝙖𝙩𝙖 𝙀𝙭𝙩𝙧𝙖𝙘𝙩𝙞𝙤𝙣 𝙛𝙧𝙤𝙢 𝙁𝙞𝙣𝙖𝙣𝙘𝙞𝙖𝙡 𝙍𝙚𝙥𝙤𝙧𝙩𝙨 𝙪𝙨𝙞𝙣𝙜 𝘿𝙊𝙉𝙐𝙏 Overview: Architected an end-to-end document understanding pipeline to extract tabular and semi-structured financial data without traditional OCR. Key Contributions: • Fine-tuned DONUT on curated financial PDFs for layout-aware interpretation. • Designed preprocessing and labeling flows for semi-structured elements including tables, notes, and cross-referenced totals. • Built training and evaluation workflows supporting variable page formats. • Integrated vision-language modeling for accurate layout-dependent extraction. • Delivered a prototype pipeline converting financial report images into structured XHTML suitable for XBRL workflows. Impact: Reduced manual data entry and demonstrated scalable, OCR-free financial data structuring using vision-language models.

Project: 𝙍𝙚𝙞𝙣𝙛𝙤𝙧𝙘𝙚𝙢𝙚𝙣𝙩 𝙇𝙚𝙖𝙧𝙣𝙞𝙣𝙜 𝙛𝙤𝙧 𝙊𝙥𝙩𝙞𝙢𝙖𝙡 𝙑𝙚𝙣𝙙𝙤𝙧 𝙋𝙖𝙮𝙢𝙚𝙣𝙩 𝙎𝙘𝙝𝙚𝙙𝙪𝙡𝙞𝙣𝙜 Built a multi-agent reinforcement learning system to optimize vendor payments under real-world budgetary and contractual constraints. Key Contributions: • Designed a MARL environment using PettingZoo to simulate interdependent vendor agents. • Created reward functions balancing discounts, penalties, and relationship management. • Modeled real-world financial decision rules to maximize cumulative savings. • Developed simulation pipelines to evaluate payment strategies across timelines. • Built an extensible framework enabling new vendors, conditions, and rules. Impact: Enabled dynamic, data-driven payment decisions aligning financial efficiency with operational realities.

Project: 𝙄𝙣𝙩𝙚𝙧𝙖𝙘𝙩𝙞𝙫𝙚 𝘽𝙤𝙤𝙠 𝙄𝙣𝙩𝙚𝙧𝙛𝙖𝙘𝙚 Designed an AI agent allowing users to converse with books, derive insights, generate question papers, and check idea originality. Key Contributions: • Built a context-aware RAG pipeline for large literary collections. • Used GCP Document AI and OCR for digitization and preprocessing. • Integrated Gemini Flash for multimodal reasoning. • Engineered prompt-chaining flows for dynamic, context-intelligent dialogue. • Implemented hybrid vector retrieval using Weaviate, Pinecone, and PGVector. • Created narrative-preserving chunking strategies for long-context understanding. Impact: Converted passive reading into interactive exploration using safe, explainable AI.

Overview: • Designed a GenAI agent powering autonomous helpdesk resolution and intelligent ticket triaging. • Authored RFP, built prototypes, and delivered successful presales pitches that secured project funding from MVP to production. • Mentored teams across Data Science, Analytics, ML, and GenAI.

Overview: • Built a recommendation engine enabling recruiters to identify top candidates for open roles. • Created monitoring dashboards for live ETL pipelines handling streaming data. • Led and mentored data science and analytics teams, specializing in recommender systems and ML best practices.

Developed demand forecasting, markdown optimization, and stockout prediction models to support e-commerce operational planning.

Built and optimized Python-based ETL pipelines for enterprise data workloads.

Developed ETL pipelines using Informatica PowerCenter and executed advanced SQL-based analytics.