Shrey Patel

Data Engineer · Observability & DevOps

• Designed and operated a closed-loop LLM evaluation platform, with automated grading pipelines on ClickHouse, Grafana, and Prometheus. Validated grader accuracy across model iterations, on multi-tenant infrastructure.
• Shipped evaluation methods into CI/CD pipelines (Great Expectations, contract tests, anomaly monitoring). Broken runs fell about 70%. Safety measurement ran continuously during model training.
• Built a high-scale telemetry generation pipeline with user-behavior simulation, for RCA workflows. The coverage gaps it surfaced informed model improvements upstream.

Data Engineer

• Stood up the multi-stage classification and ranking services (PyTorch, TensorFlow) behind abuse detection and content moderation. Optimized precision-recall to reduce false negatives on adversarial inputs.
• Designed a real-time evaluation feedback pipeline on Kafka and Flink, with a human in the loop. Harm patterns from human review became automated model evaluation signals.
• Wrote the evaluation observability dashboards (FastAPI, React, SSE, Grafana), with drill-down on model behavior across user segments at 99.9% uptime. They showed systematic benchmark gaps. Proposed red-team test sets against those gaps.

Data Engineer

• Built the real-time data processing architecture on Kafka, Flink, and Spark. It handled billions of events a day at 99.9% availability. Designed the data-quality checks in front of it, so inputs to the production scoring models stayed representative and unbiased.
• Deployed and operated production ML models on AWS SageMaker, behind end-to-end validation pipelines. Built evaluation harnesses for drift, prediction quality, and failure modes (+35% revenue, +20% CSAT).
• Expanded an Iceberg/Snowflake lakehouse to carry multimodal GenAI training data. Built lineage and validation layers under it, and latency fell about 60%. Drift-monitoring dashboards caught model regressions.

Software Engineer Intern · Data and ML Platform

• Debugged Linux production servers and analyzed network trace logs. Incident resolution time fell 18%. Wrote the readiness documents for those investigations.
• Extended the Python-based auto-detection systems. Anomaly detection speed went up 19%, and false positives went down.
• Automated the SQL and Python workflows over structured and unstructured data. Manual reporting time fell 22%.