Sravan Sama

About Sravan Sama

Machine Learning Engineer with 5+ years of experience designing, building, and deploying large-scale, production-grade ML systems across payments, risk, and real-time decisioning platforms, delivering measurable business and revenue impact. Currently at Stripe, driving global payment authorization and retry optimization using deep learning–based Adaptive Acceptance systems, improving authorization rates by 2.5–3.2% and reducing unnecessary retries by 30–35% in high-volume, mission-critical payment flows. Proven expertise in Deep Learning for tabular data (TabTransformer, TabTransformer+), PyTorch, and embedding-based models, achieving significant gains in model calibration, precision, and online decision quality under strict SLA and regulatory constraints. Strong background in real-time ML inference systems, building low-latency (<50ms P99) services using Docker, Kubernetes, REST/gRPC, and cloud-native architectures to support millions of transactions at peak scale. Hands-on experience with end-to-end MLOps, including distributed GPU training, model versioning, canary deployments, automated rollback, delayed-label learning, and continuous production refreshes for high-stakes ML systems. Deep expertise in streaming data pipelines and feature engineering, leveraging Apache Kafka, Spark (batch & structured streaming), online/offline feature stores, and strict training–serving parity to enable reliable real-time ML. Strong cloud engineering background across AWS (primary) and Microsoft Azure, scaling GPU-backed training and inference pipelines and reducing model training cycles from days to hours. Previously at Accenture, delivered nation-scale real-time ML systems serving 100M+ users, including privacy-safe risk scoring, streaming analytics, and low-latency inference during extreme traffic spikes. Experienced across classification, risk modeling, behavioral modeling, and real-time decision systems, with a strong foundation in statistics, feature engineering, imbalanced data handling, and model observability. Known for end-to-end ownership, cross-functional collaboration with product and platform teams, and translating complex ML systems into reliable, business-critical outcomes in production.

Experience

1. Machine Learning Engineer

   Stripe

   Jan 2025 — Present · Santa Clara, CA, US

   Improved global payment authorization rates by 2.5–3.2% by deploying real-time deep learning models (PyTorch) on AWSbacked cloud infrastructure, integrating low-latency inference into Stripe’s authorization and retry decision engine.• Reduced unnecessary retry attempts by 30–35% by replacing legacy XGBoost-based logic with a TabTransformer - based neural model, leveraging AWS compute and storage services to scale training and evaluation across high- volume payment traffic.• Increased retry decision precision by 60–70% by training embedding-driven models on AWS GPU instances, enabling faster experimentation and frequent production refreshes without impacting live payment reliability.• Designed and trained TabTransformer+ models in PyTorch, learning high-cardinality embeddings for issuer IDs, merchant profiles, network response codes, and time-based retry signals, improving model calibration by 20%.• Built low-latency ML inference services using Docker and Kubernetes, achieving sub-50ms P99 latency for real- time payment authorization decisions under peak traffic.• Developed Kafka-based streaming pipelines processing authorization attempts and retry outcomes at scale, reducing feature freshness lag by 40% compared to batch pipelines.• Implemented online and offline feature stores, ensuring 100% feature parity between training and inference whileeliminating data leakage and supporting delayed-label learning for declined transactions.• Scaled GPU-backed distributed training pipelines, reducing model training time from multiple days to a few hours, enabling weekly (and faster) production model refreshes.

Education

• Narsimha Reddy Engineering College

  B.Tech

• Lindsey Wilson University

  Master of Science - MS