Profile

Experienced Data Scientist with a demonstrated history of working in IT industry. Collected data and drew insights to improve business operations and solve business problems.

Professional Experience
Senior Data Scientist, Wolters Kluwer⁠
Sep 2024 – Present | Pune, India
  • •FAB (FAB Tune) as DDE/FDE
  • •Architected and delivered a multi-method AI agent evaluation platform on FAB Tune, enabling platform users to benchmark their own AI agents across quality, retrieval, and structural dimensions — deployed as independently versioned AWS Lambda microservices (sandbox → stage → prod) via SAM with OpenTelemetry distributed tracing.
  • •Built a comprehensive LLM-as-judge evaluation suite combining G-Eval (probability-weighted scoring across Consistency, Relevance, Coherence, Fluency, Usefulness), a generic runtime-configurable LLM judge (system + user prompts injected at runtime with no code changes), and a structured constraint evaluator using FHIR R5 + hierarchical JSONPath constraints with an LLM fuzzy-grader producing three-class verdicts (meets_criterion, fails_omission, fails_commission).
  • •Implemented RAG pipeline evaluation (inspired by frameworks like RAGChecker and DeepEval) with a relevancy evaluator computing precision, recall, F1, NDCG, coverage, and cumulative ranked-position P/​R curves against expert-labeled golden chunks, backed by an LLM-powered sub-document component extractor for fine-grained golden alignment.
  • •Delivered deterministic evaluation backends including ROUGE-N (1/​2/​3/​4), ROUGE-L, ROUGE-Lsum, BLEU (1–6), and a recursive JSON evaluator (leaf-level diff) producing normalized precision scores — covering both n-gram quality benchmarking and structured output correctness verification for user agents.
  • •Salesforce Case Classification
  • •Developed an AI-powered FastAPI microservice integrating Salesforce CRM and FAB (Federated Agent Backend) LLM API with optimized prompt engineering to automatically classify support cases achieving 95%+ accuracy in case reason, topic, and product area assignments, reducing manual routing time by 70%.
  • •Implemented RESTful APIs with async/​await patterns, conditional OAuth 2.0 authentication, and connection pooling for high-concurrency workloads, integrating multiple enterprise systems (Salesforce, FAB AI, Configuration APIs).
  • •Containerized and deployed the application on Azure Kubernetes Service (AKS) using Docker multi-stage builds, automated CI/​CD pipelines via Azure Container Registry (ACR), and horizontal pod autoscaling for production reliability.
  • •Borrower Analytics
  • •Developed and fine-tuned a Donut (Document Understanding Transformer) model from Hugging Face for document classification, training on PDF files to recognize 12 distinct document classes including UCC1 forms, tax liens, addendums, and legal exhibits
  • •Implemented end-to-end document AI pipeline using transformer architecture for visual document understanding, processing unstructured PDF documents without requiring OCR preprocessing
  • •Custom-trained document classifier achieving classification across specialized legal/​financial document types: STANDARD_UCC1, UCC3_ADDENDUM, TAX_LIEN, EXHIBIT, FRONT_PAGE, SUMMARY_PAGE, and other business-critical document categories
  • •Built production-ready model artifacts including custom tokenizers, configuration files, and PyTorch model weights, demonstrating expertise in transformer model fine-tuning and deployment-ready AI solutions
  • •Content Rocket
  • •Developed a cloud-native Azure Functions application using FastAPI and Python that processes documents through OCR and AI analysis, integrating Azure Cognitive Services Computer Vision API for text extraction and Azure OpenAI GPT-4 for intelligent content generation
  • •Built a robust document workflow system that accepts PDF and text uploads, performs multi-page OCR with configurable page ranges, and generates structured responses based on custom prompts and user instructions with comprehensive error handling and logging
  • •Engineered scalable AI integration using the Instructor library for structured LLM responses, Pydantic models for data validation, and Jinja2 templating for dynamic prompt management, enabling flexible document analysis and content generation workflows
  • •Business License Management System (BLMS)
  • •Developed an AI-powered business license recommendation system using Python, FastAPI, and Azure OpenAI GPT models to analyze business descriptions and suggest relevant licensing requirements across multiple industries
  • •Implemented intelligent query processing with multiple AI approaches including knowledge-base search, taxonomy-based classification, and GPT-powered natural language understanding to match user queries with appropriate license categories
  • •Built scalable microservice architecture with Docker containerization, Kubernetes deployment configurations, and environment-specific overlays (dev/​qa/​prod) supporting caching mechanisms and embedded vector data processing
    • Senior Data Scientist, Arya.ai⁠
    Dec 2022 – Aug 2024 | Remote, India
  • •Bank Statement Analyser(BSA): ⁠
  • •Using computer vision techniques and algorithms such as image processing, pattern recognition, and optical character recognition (OCR), I developed a solution that improved the accuracy of detecting transactions and other details in bank statements by over 95%. (API link⁠)
  • •The implementation of the computer vision solution had a significant impact on the organization, reducing manual labor, improving accuracy, and saving time.
  • •Deployed and managed a RESTful API on a Google Cloud Platform using Docker for 20+ countries, ensuring scalability, security, and optimized performance.
  • •Collaborate closely with clients to understand their unique objectives, translating their requirements into actionable steps for the implementation of our machine learning solution, leading to highly customized and effective deployments.
  • •Bank Statement Extraction(BSE): ⁠
  • •Built an API for extracting basic details like account holder name, A/​C number, statement duration, etc.
  • •Used OpenAI GPT model to get these basic details from raw OCR output.
  • •Successfully deployed the API on Google Cloud Platform (GCP), ensuring scalability and reliability for seamless integration with the application's backend infrastructure.
  • •Intelligent Document Retrieval and Chatbot System:
  • •Document Processing:
  • •Implemented efficient chunking logic to partition documents into manageable segments for streamlined processing.
  • •Conducted rigorous testing of chunking efficiency, fine-tuning parameters to optimize performance.
  • •Model Deployment and Integration:
  • •Selected and deployed embedding and Large Language Model (LLM) models compatible with AWS Bedrock, ensuring seamless integration with the system architecture.
  • •Configured and deployed ChromaDB on AWS as a client-server, establishing Lambda triggers for real-time document updates in S3 and efficient embedding addition/​update operations
  • •Document Access Control and Chat History Management:
  • •Designed and implemented a schema for storing chat history in a relational database (RDS), capturing and storing each conversation with precision.
  • •Chatbot Pipeline Development:
  • •Developed a sophisticated pipeline to process user questions, user IDs/​usernames, and chat history seamlessly.
  • •Integrated logic to determine document access based on user ID, delivering tailored responses accordingly.
  • •Monitoring and Validation:
  • •Established CloudWatch metrics for monitoring Lambda calls and outputs, ensuring system reliability and performance.
  • •Conducted extensive validation and testing of the end-to-end pipeline, crafting benchmark questions and meticulously assessing retrieval and response accuracy.
  • •Documented testing outcomes comprehensively, iteratively refining the system for enhanced performance.
  • •API Gateway Setup:
  • •Orchestrated the setup of Lambda functions and API gateway to facilitate seamless API interactions, enabling effortless integration with external systems.
    • Oct 2021 – Dec 2022 | Pune, India
  • •Developed a machine learning model using Python and scikit-learn library that predicted the likelihood of insurance claims being fraudulent based on claimant information and historical claims data.
  • •Evaluated the model's performance using metrics such as precision, recall, and F1-score, and presented the results to stakeholders.
  • •The project resulted in a 20% reduction in fraudulent claims, saving the insurance company millions of dollars in losses.
  • •Azure cloud migration- Migrating on-prem (ASIS) to Azure cloud (ToBe)
    • Dec 2018 – Oct 2021 | Pune, India
  • •Trained classification model to classify alerts received for different application via internal monitoring portal and triggering email notifications based on output class.
  • •Getting alerts insights using pandas and NumPy- Scraping the data using selenium and BS4 to get the raw data from alert reporting tools such as WebGui and analyzing this data to avoid unnecessary failures
  • •Monitoring Informatica, Abinitio and Bigdata jobs through TWS, SOA and Informatica
  • •Triggering automatic emails to client for alerting visa batch start and end by reading live SVG graphs/​ pictures from website using python
  • •Generating daily reports such as Bigdata (Hound cluster and Omega cluster) job completion using TWS scrapped data with selenium and BS4
    • Skills
    Agentic AI

    LangGraph, Smolagents, Crewai, Pydantic AI

    Deep Learning

    DNN, CNN, RNN, Transfer learning, LSTM, LLM, GenAI, RAG

    Programming Languages and tools

    Python, Java, HTML and CSS

    Python Libraries

    Sklearn, Scipy, Statsmodel, Pandas, NumPy, Seaborn, Matplotlib, Selenium, BS4, NLTK, TensorFlow, Keras, pytorch, langchain, openai, hugginface, shap, Boto3, Flask, Fastapi, Streamlit

    Database Language

    SQL

    Cloud Tools

    AWS (sagemaker, s3, Lex, Lamda, Bedrock, RDS, EC2, API gateway, secret manager), GCP (Bigquery,

    Dataproc, Dataflow, Vertex AI, cloud SQL, etc.)

    Machine Learning

    Classification, Regression, clustering, Decision Trees, K-Means Clustering, hierarchical clustering

    Statistical Methods

    Predictive analysis, Hypothesis Testing and Confidence Intervals, Principal Component Analysis, LDA and Dimensionality Reduction

    Version Control Tools

    Git

    Scripting Language

    Unix

    Data Reporting Tool

    Excel, GCP Data Studio

    No Code Tools

    N8N, Make.com, Zapier

    Personal Projects
    WeldCraft, Hybrid Search, RAG
    Apr 2024 – Apr 2024
  • •Deployed Chroma (an open-source vector database) on EC2 with a security layer.
  • •Developed a hybrid search (keyword search + semantic search) algorithm for pdfs and web page using langchain
  • • Built RAG chatbot with history for these pdfs and web pages using langchain and openai(gpt-3.5-turbo)
  • •Deployed hybrid search and RAG chatbot as Fastapi endpoint and nginx
  • •Designed a webapp for accessing these endpoints using nextjs, tailwindcss.
  • •Built CI/​CD pipeline using GitHub actions and ec2 as self-hosted runner.

    • Tech: Langchain, Vector databse, RAG, Openai, llm, github actions, ec2

    Nov 2023 – May 2024
  • •Designed and implemented a dynamic document-based Q&A platform, enabling users to upload diverse file formats (PDFs, Excel, JSON, text) for intelligent querying.
  • •Integrated OpenAI and LangChain for advanced natural language processing, while securely storing documents on AWS S3.
  • •Implemented Clerk for robust authentication, Pinecone for efficient document indexing, and PostgreSQL (Neon DB) for seamless chat and message history storage.
  • •Additionally, incorporated Stripe for streamlined payment processes, enhancing the overall user experience.
  • •Crafted a responsive and intuitive front-end experience using Next.js, TypeScript, and Tailwind CSS, complemented by a robust Node.js backend. Implemented cutting-edge technologies to ensure a seamless user interface, blending the power of Next.js for efficient rendering, TypeScript for type safety, and Tailwind CSS for a sleek and modern design. The result is a dynamic and user-friendly platform that enhances the overall usability and accessibility of the application.
  • •Deployed the webapp on vercel.

    • Tech: OpanAI, RAG, Langchain, Pinecone, AWS S3, NextJs, Typescript, Tailwindcss

    Harmony- Extracting Questionnaire⁠, Machine Learning Classification, NER, Open Source
    Feb 2024 – May 2024
  • •Aim was to extract questionnaire from given text and Json input without using spacy.
  • •Developed a robust solution to extract questionnaires from unstructured text and JSON inputs using advanced NLP techniques.
  • •Conducted thorough preprocessing of both text and JSON data to ensure optimal performance of the models.
  • •Implemented two classification models:
  • •XGBoost classifier for JSON data processing.
  • •LSTM for text data analysis.
  • •Employed fine-tuning techniques on BERT for Named Entity Recognition (NER) to eliminate irrelevant information from the output of classification models.
  • •Achieved enhanced accuracy and efficiency in questionnaire extraction through iterative model training and refinement.

    • Tech: XGBoost, LSTM, Pytorch, TensorFlow, HuggingFace

    Certificates
    Education
    Data Science and Engineering, Great Lakes⁠
    Apr 2021 – Apr 2022
    Jul 2014 – Jul 2018