• Engineered a Retrieval Model for AI Agents from scratch with a novel in-context chunking strategy — boosting relevance across long-context queries.

• Designed a custom algorithm to extract key–value pairs from unstructured documents, which now powers real-time quoting for Insurance contracts.

• Fine-tuned a LLaMA model to generate structured responses autonomously — no prompting required — using minimal supervision and internal feedback loops.

• Pioneered evaluation of LLMs using logit-based statistical methods, integrating techniques like LLM-as-a-Judge, Rank Consistency, and Response Entropy to benchmark hallucination rates and factual grounding.

• Always experimenting with tools like RAG, DeepEval, phi-3, and vector search hybrids to push the boundaries of intelligent agent capabilities.

• Engineered a high-throughput OCR pipeline using LayoutLMV3 and PyTorch Lightning, optimized for multi-GPU training to process massive document datasets while managing memory constraints.

• Architected a scalable data ingestion system with Apache Kafka, setting up new topics and brokers to handle high-volume, real-time data streams for the OCR model.

• Fine-tuned a custom language model to perform post-OCR validation and correction, automatically identifying and rectifying inaccuracies to ensure high-fidelity data extraction.

• Engineered a multi-modal deep learning framework to detect subtle instances of racial bias in text, speech, and video content. The model utilized a hybrid attention mechanism with Bi-LSTMs and Transformer encoders, achieving 87% accuracy while minimizing false positives on a highly nuanced, custom-built dataset.

• Integrated facial emotion recognition using OpenCV as a key feature for the multi-modal bias detection system, improving context understanding in video analysis.

• Implemented a few-shot learning pipeline with GPT-3 and multi-threaded calls to the OpenAI Moderation API to establish a robust baseline and augment training data for hard-to-classify examples of implicit bias.

• Collaborated at Transfer Scheduler, facilitating transfer of terabytes of data from Google Drive, S3using Kafka and RabbitMQ.

• Implemented PKI secrets engine using Vault, enabling token management with fine grained ACL policies..

• Developed an NLP model to optimally convert code from one syntax to another, finding ambiguities and resolution, Achieved an accuracy of 88% for Cobol to Java for over 10 million lines of code.

• Invented a script to map and create all different variables, removed multiple duplicated variables in project code.

• Migrated Western Union's money transfer solution from Cobol to Spring Java Microservice, re-designed and designed architecture of WillCallStore (WCSTOX) program's solution by collaborating with 10+ teams, transactions from 50+ countries.

• Developed WCSTOX Program from scratch, validates and stores all money order records in postgres and connecting data-stream Kafka, managing configurations of multiple Kafka topics, Devised common encryption layer for all user sensitive data.

• Steered team of 10 new developers and resolved task of stitching code piece by piece, reporting to higher management.