Gowri Kajipuram

About

ML/LLM engineer who has built and productionized a large multimodal LLM pipeline end-to-end—fine-tuning a 20B+ parameter model with distributed/FSDP training and deploying on Kubernetes via Triton for ~5x throughput. Strong focus on reliability and safety (monitoring with SHAP, guardrails, A/B testing) with reported ~22% relevance lift and reduced harmful/incorrect outputs, plus experience orchestrating ETL/retraining workflows with Airflow across S3/Snowflake/RDS.

Experience

Education

Key Strengths

• Built and deployed multimodal LLM system to reduce hallucinations and improve image+text understanding
• Scaled training of 20B+ parameter model using distributed training + FSDP
• Improved training efficiency via precision tuning and gradient checkpointing (~40% time reduction)
• Productionized LLM serving on Kubernetes with Triton (≈5x throughput under peak load)
• Implemented monitoring and safety controls (e.g., SHAP) with measurable quality gains (≈22% relevance lift, reduced harmful/incorrect outputs)
• Designed structured LLM/agent evaluation with KPIs (latency, hallucination rate, task success, token efficiency) and large-scale A/B testing
• Built end-to-end orchestration for data/ML lifecycle (scheduled ETL, dependencies, automated retraining) across S3/Snowflake/RDS
• Effective collaboration with non-technical stakeholders; delivered interpretable outputs via Power BI dashboards