• Led cloud-native deployment of distributed systems on AWS and GCP, leveraging Kubernetes (EKS/GKE) to build scalable, highly available services across multi-region environments using Terraform.

• Built and optimized event-driven data pipelines using Kafka and RabbitMQ, processing 500M+ messages/day with S3/ADLS backed storage, reducing event propagation latency by 70%.

• Developed observability and monitoring solutions using Grafana, ELK, CloudWatch, and GCP Profiler, improving system visibility and reducing MTTR by 35% across production services.

• Built and deployed LLM inference pipelines using vLLM and Triton Inference Server, improving system scalability and resource utilization across Kubernetes-based deployments.

• Developed end-to-end LoRA and SFT fine-tuning pipelines for domain-specific models, improving compliance text classification accuracy using BigQuery and Snowflake feature engineering workflows.

• Designed and deployed FastAPI-based ML services serving 1M+ daily requests, leveraging async processing and Redis caching to reduce p95 latency by 60%.

• Designed and deployed identity resolution microservices in Python/C++, unifying 50M+ Entra ID and HR records with versioned audit trails, self-healing workflows, and high availability.

• Refactored a monolithic system into scalable microservices architecture, implementing contract testing and phased rollouts to improve system throughput by 45% without regressions.

• Built high-throughput data pipelines using Kafka and Spark Streaming, processing 200M+ events/day with secure archival to S3/Glacier and compliance with FINRA/SEC retention policies.

• Built and deployed LLM inference services using Triton Inference Server and vLLM, enabling scalable, production-grade model serving for enterprise workloads.

• Optimized large language model performance (GPT, BERT, LLaMA) through quantization and efficient inference strategies, reducing memory usage by 60% and improving latency.

• Developed and optimized multi-modal inference pipelines with advanced decoding strategies (Greedy, Beam, Sampling), improving token generation performance by 20%.