Nithin Jella

Architected an enterprise multimodal RAG pipeline using LlamaIndex with Qdrant vector search and a quantized Llama-3-8B model, retrieving context from 4TB+ internal documentation to automate customer support resolution by 34%. Optimized deep learning inference by compiling PyTorch models to TensorRT and deploying via Triton Inference Server on AWS EKS, reaching 4,500+ QPS and reducing P95 latency from 850ms to 120ms for real-time NLP. Built distributed data ingestion with PySpark and Delta Lake on Databricks for daily telemetry streams (50M+ events), feeding curated features into Feast feature store for downstream model consumption. Executed distributed fine-tuning using DeepSpeed, QLoRA, and Hugging Face across multi-GPU clusters to improve domain-specific entity extraction accuracy while reducing cloud compute cost. Developed a multi-agent orchestration layer using LangGraph and Python microservices to automate dynamic, context-aware workflows by chaining external APIs and enterprise

Developed gradient-boosted risk classification models using LightGBM and scikit-learn, engineering localized financial features to improve baseline insurance underwriting prediction accuracy by 14%. Configured end-to-end model tracking and lifecycle lineage using MLflow and Weights & Biases, enabling structured hyperparameter optimization sweeps that accelerated experimental model convergence. Built automated ELT data ingestion workflows with Python and dbt to clean inconsistent Snowflake data lake tables, eliminating critical null-value anomalies and reducing downstream reporting latency by 28%.

Designed and trained custom BERT-based text classification models to categorize inbound customer service tickets, automating triage workflows and diverting 42% of tier-1 support requests to appropriate resolution queues. Built a real-time, scalable recommendation engine using collaborative filtering with scikit-learn, exposed via highly concurrent FastAPI microservices handling up to 1,200 API calls per minute during peak hours. Orchestrated robust asynchronous data streaming pipelines using Apache Kafka and Apache Airflow to ingest and aggregate continuous transaction data, ensuring data freshness and consistency for downstream predictive analytics models. Containerized complex ML environments with Docker and orchestrated deployments on Azure Kubernetes Service (AKS), optimizing pod resource allocation to support business-critical inference with 99.9% uptime. Conducted A/B testing and statistical analysis to evaluate champion/challenger model performance, supporting rollout of an upgr