Data Scientist | Generative AI | Machine Learning | AI Strategy Consultant

projects

hashnode

github

Shreyas Kulkarni  • Peerlist

<ul><li>Led the design and end-to-end development of a hybrid query routing system as a proof of concept (POC) for a major US-based auto insurance client, which directly contributed to securing the project.</li><li>Built a LangChain-powered data pipeline using Qwen-QWQ-32B on Groq for dynamic SQL query generation and retrieval from Postgres, enabling accurate, LLM-generated responses to user queries.</li><li>Developed a RAG workflow with ChromaDB and Google’s Text Embedding 004 for vector-based retrieval, integrating UnstructuredLoader to process PDF policy documents and deliver context-rich answers to policy-related questions.</li><li>Reduced hallucinations and improved trust in AI outputs by integrating a hallucination grading and reranking mechanism using cross-encoders, significantly enhancing response quality during client demos.</li><li>Engineered an intelligent hybrid router that dynamically split user queries across relational and vector stores, optimizing for speed and accuracy; showcased as a key differentiator in the sales demo.</li><li>Collaborated with a sales-aligned SME for domain guidance and worked alongside a frontend developer for UI integration, ensuring a complete and demo-ready POC.</li></ul>

Intelligent Query Routing and Answering System

<ul><li>Developed a GenAI-powered LLM module using LangChain, integrated with Llama 3.1 (8B) via Ollama, to automate the extraction and structuring of employee information from resumes, reducing HR lookup time by 80% and improving HR team efficiency in candidate evaluation.</li><li>Developed the Top 5 Skills Identification Module using a Llama 3.1-based solution, which helped increase skill categorization accuracy by 89% over manual keyword-based methods, enabling HR teams to better assess candidate fit for roles.</li><li>Led architectural decisions within the scope of LLM-related modules with LangChain, focusing on resume extraction and skill identification</li><li>Utilized hybrid prompting techniques (Instruction-based, Constraint-based, Structured JSON Output) for extracting skills, experience, and project details from unstructured resumes, ensuring structured, reliable output in the required JSON format.</li><li>Collaborated closely with HR teams to understand their requirements and refine the resume extraction and skills identification process. Worked cross-functionally with data analysts and Python developers to ensure seamless integration of LLM-based modules.</li></ul>

Automated Employee Profiling & Resume Processing Pipeline

Progressed from Junior to Data Scientist over a 3+ year tenure, taking ownership of core modules and mentoring new team members and remained a core contributor across all project phases.<ul><li>Built and scaled a robust document parsing pipeline for healthcare PDFs with highly inconsistent layouts, powering a live database of 800,000+ US medical and pharmaceutical formulary documents using Python, Pandas and table extraction libraries like Tabula and Camelot</li><li>Enabled clients to launch 3 commercial subscription-based B2B products that help stakeholders track coverage changes and analyze formulary status across health plans using outputs from this pipeline.</li><li>Designed and implemented a deep learning–based document routing system using Vision Transformers (ViT) and CNNs replacing rule-based extractors achieving 92% routing accuracy and reducing downstream processing failures by 50%.</li><li>Led fine-tuning and experimentation of ViT models (via PyTorch and Transformers) for layout classification and table presence detection across diverse document templates.</li><li>Replaced unreliable OCR/table extractors by integrating Detectron2 (Mask R-CNN) to predict column lines, drawing them programmatically using PyMuPDF, and extracting data using both Camelot and Tabula, with quality-based selection logic to pick the better output.</li><li>Fine-tuned custom spaCy NER models to extract clinical entities like dosage, drug names, and restrictions from noisy policy text, enhancing downstream entity resolution performance.</li><li>Effectively used Git to manage a rapidly evolving codebase, submitting over 500 PRs across 3+ years; remained the sole project constant through changing team sizes (2–6), consistently delivering on client-specific requirements.</li><li>Mentored all new team members, onboarding them into a large and evolving codebase, providing KT sessions and driving best practices in model development, data cleaning, and debugging.</li></ul>

Automated Formulary Parser for Healthcare Documents

Collaborating with the authors to ensure technical accuracy and providing valuable insights to enhance the book's content.