• Automated Incremental Data Ingestion for Search with a single centralized Self-service microservice that seamlessly handles concurrent requests from multiple clients. This helped significantly reduce PoC turnaround times and developer involvement.• Reduced time taken for table extraction from PDFs by 85% using Multiprocessing in Python.• Enhanced Solr retrieval for Search by augmenting user queries and indexed records with hierarchical or non-hierarchical metadata extracted from raw client data and enabling metadata-based search. Automated metadata extraction and tagging of indexed records using Cron jobs.• Leveraged LLMs for RAG-based generation grounded in client Knowledge-Base data and for summarization of complex/verbose user queries.• Generalized and integrated client-specific Structured Data Ingestion and Search modules as part of mainstream pipelines with suitable feature flags and configurable parameters.• Leveraged Google’s Document AI for OCR on scanned PDFs, extracting tables and text with near-perfect accuracy. Also performed Form Field extraction to answer objective queries from PDFs.• Developed a modular Python Pipeline with unit tests to perform grouped Agglomerative Clustering using Scikit-Learn with finetuned Sentence Transformer embeddings for deduplication of predictor-specific observation and resolution texts by directly fetching data from Solr cores. Improved the average cluster cosine similarity by 60% and significantly reduced the runtime from many hours to a few minutes.• Deployed Flask and FastAPI endpoints as services on Azure and GCP Linux environments.• Collaborated effectively with Cross-functional teams for Client Project Deliveries, Bug Resolutions, and Production Deployments. Led end-to-end Search implementation for 2 major clients, while contributing to multiple other successful PoCs.

1. Developed an Open Domain Question Answering System using Haystack Framework, based on Dense Retrieval. Uses the FAISS Document Store and Dual Encoder for queries and passages. Pretrained transformer models from HuggingFace were used for the encoders. Integrated a regular Closed Domain QnA system as well in the same project. 2. Developed an automatic answer evaluation system for factual and descriptive answers. Finetuned a DistilBert model on TREC dataset for classifying questions into above categories. The former uses NER and subject-object dependency for evaluation, while the latter employs a combination of Sentence-Bert cosine similarity and BLEU score between the reference and candidate answers.Both above projects developed on PyCharm IDE; also created Streamlit demos as PoC.3. Performed classification of various room images into 6 distinct categories by finetuning a MobileNet-V2 model on the ADE20k dataset (sourced from Kaggle).Prepared detailed documentation for all projects.

Development of Operation Support System and User Interface for the Robotic Optical Management Engine (ROME), for remote access via SSHClient using Django Framework and SQLite DB backend. The common interface handles multiple users and commands the engine to execute various operations based on user input through the GUI. Received a PPO.

Problem Statement involved binary text classification, automatic headline generation and Named entity recognition of tweets and articles about Mobile Tech companies. Our Team finished with a bronze in the event.