DIVAKAR PANDEY

Senior Software Engineer with 6+ years of experience designing scalable backend systems and AI-powered enterprise applications within the financial domain. Specialized in Python, FastAPI, Large Language Models (LLMs), NLP, Computer Vision, Document Intelligence, Semantic Search, and Microservices. Experienced in building production-grade AI solutions using Model Context Protocol (MCP), Vector Databases (pgvector), YOLOv8, PaddleOCR, HuggingFace, Neo4j, Elasticsearch, SQLAlchemy, Docker, and Celery. Proven track record of leading AI initiatives from research and proof-of-concept through production deployment while mentoring engineers and collaborating with cross-functional teams.

• IBM Data Science Specialization (Machine Learning, Deep Learning, TensorFlow, PyTorch)
• Neural Networks and Deep Learning – Andrew Ng (Coursera)

• Delivered multiple AI and backend platforms from proof-of-concept to enterprise production environments.
• Improved enterprise table detection accuracy from 70–80% to 97% through YOLOv8 fine-tuning.
• Built semantic retrieval systems leveraging vector embeddings, pgvector, and Large Language Models.
• Led deployment and technical delivery of AI initiatives while mentoring junior engineers.
• Designed scalable document-processing pipelines capable of handling millions of financial documents.

Enterprise Guidance Extraction Platform

Python | FastAPI | LLM | MCP | pgvector | PostgreSQL

• Designed an LLM-powered platform to extract structured financial guidance from earnings reports and SEC filings.
• Built custom Model Context Protocol (MCP) tools enabling AI agents to answer company-specific business questions using enterprise financial data.
• Implemented semantic retrieval using pgvector to map extracted guidance with financial line items through vector similarity search.
• Engineered prompt templates and LLM workflows to generate structured outputs with improved accuracy.
• Integrated the platform into enterprise financial workflows supporting production-scale document processing.

AI-Powered PDF Table Extraction Platform

Python | YOLOv8 | PaddleOCR | Camelot | Adobe Extract API | AWS Textract | Celery

• Redesigned the enterprise table extraction engine by fine-tuning YOLOv8 on custom financial datasets.
• Improved table detection accuracy from approximately 70–80% to 97% through custom annotation and model optimization.
• Built automated annotation pipelines using Camelot, Adobe PDF Extract API, OCR, and AWS Textract.
• Integrated PaddleOCR for multilingual scanned document processing supporting both English and non-English financial reports.
• Developed scalable asynchronous document-processing pipelines using Celery for enterprise production workloads.

Deep Semantic Mapping

Python | BGE-1.5 | HuggingFace | FastAPI | pgvector

• Designed an embedding-based semantic mapping platform for financial line-item understanding.
• Generated vector embeddings using BGE-1.5 and mapped extracted financial line items to standardized financial concepts using cosine similarity.
• Replaced keyword-based matching with semantic retrieval to significantly improve mapping accuracy.
• Developed reusable embedding services supporting multiple downstream AI applications.

Enterprise Knowledge Platform using MCP

Python | FastAPI | MCP | PostgreSQL | pgvector | LLM

• Designed and implemented a custom Model Context Protocol (MCP) server for enterprise AI assistants.
• Developed reusable MCP tools enabling contextual question answering over structured and unstructured company data.
• Built vector search architecture using pgvector supporting semantic retrieval and alias search.
• Exposed scalable APIs enabling seamless integration with enterprise chatbot applications.

Enterprise Search Platform

Python | Elasticsearch | Kibana | FastAPI

• Developed an enterprise search platform indexing 4.5+ million financial research documents.
• Implemented fuzzy search, autocomplete, synonym search, BM25 ranking, and optimized Elasticsearch queries.
• Built Kibana dashboards for operational monitoring and analytics.
• Improved search relevance using edit-distance algorithms and semantic similarity techniques.