Sumanth Dola

• Designed and scaled end-to-end LLM training pipelines to support rapid experimentation and large-scale model training, reducing end-to-end experiment setup time by 30% for research teams. • Built distributed data ingestion and preprocessing pipelines to curate and validate multi-terabyte training datasets, improving data quality checks and reducing training restarts by 25%. • Implemented training orchestration and job scheduling systems to optimize GPU utilization across clusters, increasing effective compute efficiency by 20% while maintaining fault tolerance. • Integrated experiment tracking, metadata logging, and model versioning into the training workflow, enabling reproducible research and reducing duplicated experimentation by 35%. • Collaborated closely with research scientists to productionize experimental training code into robust, restartable pipelines with automated checkpointing and recovery. • Developed training-time evaluation hooks to monitor loss curves, convergence behavior, and early regression signals, shortening iteration cycles by 15%. • Optimized dataset sharding and sampling strategies to improve training stability and convergence consistency across large-scale runs. • Partnered with infrastructure and reliability teams to enforce security, access control, and compliance standards across training environments, reducing pipeline-related incidents. • Authored internal documentation and operational runbooks for training infrastructure, improving onboarding speed and self-service adoption for new ML engineers and researchers.

• Designed and built shared ML platform components to support model development and deployment across payments, fraud, and risk teams, improving model on boarding speed by 30%. Developed a centralized feature store using batch and streaming data pipelines, enabling consistent feature reuse and reducing training–serving skew incidents by 40%. • Implemented scalable data validation and feature quality checks to detect missing, stale, or anomalous signals, improving downstream model reliability and trust. • Built automated training and retraining pipelines for multiple production models, reducing manual intervention and shortening model refresh cycles by 35%. • Integrated model versioning, experiment tracking, and metadata logging into the platform, enabling reproducibility and faster root-cause analysis of model performance changes. • Partnered with fraud, payments, and risk teams to standardize feature definitions and modeling workflows, improving cross-team collaboration and reducing duplicate feature engineering effort. • Supported low-latency online feature serving for real-time decisioning use cases, meeting strict SLA requirements for transaction processing. • Collaborated with infrastructure and SRE teams to improve platform reliability, monitoring, and alerting, reducing ML pipeline-related incidents by 25%. • Authored platform documentation and onboarding guides, increasing self-service adoption of ML tooling across engineering teams.