Dhruv Donga

Built a production-grade BERT-based NLP text-classification pipeline with Hugging Face Transformers and PyTorch to process 50,000 academic documents with sub-100ms inference latency. Developed graph neural network (GNN) models and utilities to map citation relationships across research domains, uncovering knowledge clusters for grant proposal preparation. Designed and deployed a scalable FastAPI backend on AWS EC2 with Docker, exposing LLM inference endpoints to multiple research teams at sub-200ms p99 latency and automated CI/CD via GitHub Actions to reduce manual deployment time. Evaluated multiple frontier LLMs using LangChain evaluation pipelines and RAG experiments to produce cost-saving comparative insights for API licensing.

Developed an LSTM-based NLP sentiment-analysis engine in TensorFlow for a Fortune 500 retail client, classifying millions of feedback records monthly and reducing misrouted escalations. Architected an end-to-end MLOps pipeline on AWS SageMaker with MLflow model versioning and GitHub Actions CI/CD gates, cutting deployment cycles across multiple production environments. Optimized Apache Spark ETL jobs by restructuring shuffle stages to significantly reduce runtime and AWS EMR compute costs. Containerized microservices on Amazon EKS with Kubernetes and blue-green deployments to achieve zero-downtime releases and faster recovery. Built a RAG system using LangChain and FAISS/Pinecone-style vector search components with low-latency responses for an internal knowledge-base chatbot. Migrated batch inference to an ONNX Runtime-optimized AWS Lambda pipeline using quantization to improve throughput and reduce memory footprint.

Designed production RESTful API endpoints in Spring Boot for a tier-1 banking client’s mobile application, supporting OAuth 2.0 authentication, encryption, and high uptime with low median latency. Migrated a legacy monolith to an event-driven microservices architecture on AWS using Kafka, reducing deployment time and monthly infrastructure costs. Integrated unified CI/CD pipelines using Jenkins and GitHub Actions to improve test automation and reduce manual release coordination. Built a PostgreSQL data model with read-replica sharding and PgBouncer pooling to support high concurrency, paired with Redis caching to reduce query latency. Refactored a React frontend into deployable micro-frontends using Module Federation, improving page load time and reducing duplicated UI code. Automated AWS environment provisioning with Terraform to reduce setup time and prevent production issues.