SAIRAM KOVA

•⁠ ⁠Developed Python-based data ingestion and processing workflows for Vectis, a security data analytics platform, working with AWS S3, PostgreSQL, DuckDB, and Parquet to support analysis of audit and security event data.

•⁠ ⁠Built a proof of concept for incremental ingestion of PostgreSQL AuditDB records, implementing polling, checkpoint management, retry handling, and duplicate prevention to support reliable processing across multiple AWS accounts.

•⁠ ⁠Implemented embedding generation and vector similarity search using Sentence Transformers and DuckDB VSS, allowing related records and graph entities to be retrieved based on semantic similarity.

•⁠ ⁠Integrated Graph V2 edge data from Parquet files in Amazon S3 into embedding pipeline, ensuring schema validation, efficient batch processing, incremental indexing, and accurate record identification.

•⁠ ⁠Improved the embedding workflow by adding content-based change detection, checkpointing, and duplicate handling to avoid unnecessary reprocessing of unchanged records.

•⁠ ⁠Developed analytical features leveraging cosine similarity, baseline comparisons, and statistical scoring to assess security activity pattern changes and detect potential anomalies. •⁠ ⁠Contributed to the AI Commons Model Repository by working on Hugging Face model upload and download functionality, S3-compatible storage integration, and backend API improvements.

•⁠ ⁠Tested model repository functionality with MinIO and S3-compatible storage, addressed code review feedback, and worked with engineering teams to validate API behavior and storage operations.

•⁠ ⁠Collaborated with security, infrastructure, and engineering teams on technical discussions, pull request reviews, automated testing, and integration planning for Vectis and AI Commons services.

• Designed and deployed machine learning models for credit risk assessment and fraud detection, leveraging Python, TensorFlow, and Scikit-learn to improve detection accuracy by over 15% while reducing false positives.
• Built scalable data pipelines using Apache Spark and Airflow to process multi-terabyte financial transaction datasets, cutting model training latency by 40% and improving real-time data availability for analytics.
• Implemented NLP-based models using BERT and Hugging Face Transformers to extract insights from unstructured financial documents, enabling automated compliance reporting and reducing manual review time by 30%.
• Integrated MLOps workflows with MLflow and Kubernetes for continuous model retraining and deployment across AWS SageMaker and Azure ML environments, ensuring high availability and audit compliance.
• Collaborated with data scientists and business analysts in Agile sprints to translate banking use cases into machine learning solutions, leveraging CI/CD pipelines and Docker for streamlined experimentation and delivery.
• Created interactive data visualization dashboards in Tableau and Power BI to monitor model performance, drift metrics, and key financial indicators, driving data-driven decision-making across risk and product teams.

• Built and fine-tuned deep learning models, including CNNs and RNNs, for computer vision and NLP applications, improving document classification accuracy by 25% and reducing manual review effort.
• Engineered and deployed transformer-based NLP pipelines using BERT and Hugging Face Transformers for sentiment analysis and text summarization, boosting content moderation efficiency.
• Developed data ingestion and ETL workflows with Apache Kafka and Airflow, enabling real-time data streaming and processing for analytics and model training pipelines.
• Designed and maintained SQL and NoSQL databases (MySQL, MongoDB, Redis) to support large-scale data storage and retrieval for high-throughput ML workloads.
• Accelerated model training workflows using GPU-enabled TensorFlow and PyTorch, achieving 3x faster training times and enabling iterative experimentation.
• Built and automated model monitoring dashboards with Matplotlib and Seaborn, providing actionable insights into drift, accuracy decay, and data quality issues.
• Collaborated with product and engineering teams to convert prototype ML solutions into production-grade APIs, following Waterfall-to-Agile migration for faster product iteration.