Sourabh Desai

Micro-Service for Document Deduplication

• Developed a microservice for document deduplication using Locality-Sensitive Hashing (LSH) to efficiently identify duplicate records.
• Integrated MongoDB and Azure Blob Storage to fetch and process records for deduplication.
• Designed and implemented asynchronous REST APIs to provide real-time document duplication responses.
• Coordinated with the Software Engineering and Platform Engineering teams to ensure seamless integration and deployment.
• Evaluated the deduplication service, achieving a 94% recall, ensuring high accuracy in identifying duplicate documents and minimizing false negatives.
• Deployed the microservice on Rancher, developing Helm charts and configurations, and integrated CI/CD pipelines for automated deployment.
• Tested the solution across multiple environments, including Dev, QA, UAT, and Prod, ensuring stability and performance.

Document AI Extraction

• Built and optimized Doc-AI pipelines to extract structured information from PDFs and Excel files based on auditors requirements.
• Automated data ingestion workflows to store extracted insights into an SQL database for further analysis.
• Developed regex-based extraction techniques for parsing key fields from unstructured PDFs, improving data accuracy and consistency.
• Built an automated Excel extraction pipeline, handling multiple formats, applying preprocessing techniques, and standardizing data for downstream processing.
• Enhanced document extraction using LLM models, improving text parsing, entity recognition, and context understanding for complex documents.
• Optimized and debugged Python-based workflows, ensuring efficient data extraction, error handling, and performance improvements in processing pipelines.

LLM – Powered Research analyst Q&A Chabot

• Designed and developed an AI-powered research assistant by collaborating with the client to define process flows and optimize information retrieval.
• Implemented LangChain’s UnstructuredUrlLoaders to fetch online data and utilized recursive text splitters to structure the extracted content efficiently.
• Leveraged OpenAI and Claude (Anthropic) embeddings with Pinecone for vector-based semantic search, enabling accurate and efficient document retrieval.
• Integrated OpenAI’s and Claude’s LLMs with a retrieval system to generate contextually rich and comprehensive responses.

E – Commerce Sentiment Analysis system

• Developed an NLP-driven sentiment analysis model by collaborating with the client to understand and preprocess customer review data.
• Implemented a text preprocessing pipeline including stemming, stop-word removal, tokenization, and padding to enhance data quality for modeling.
• Built and trained a Bidirectional LSTM model using Word2Vec embeddings, achieving an 85% F1-score, a 12% improvement over the baseline.