I am a computer science researcher and systems engineer specializing in the optimization of distributed cloud systems and resource-efficient machine learning infrastructure. My work focuses on the intersection of streaming data architectures and real-time predictive modeling, specifically investigating the tradeoffs between algorithmic latency, computational cost, and model calibration reliability.

An investigation into real-time predictive probability frameworks to mitigate digital alert fatigue and minimize transactional LLM compute overhead.

• The Research Problem: Traditional notification systems rely on static, rule-based heuristics that fail to adapt to live human behavioral shifts, causing systemic cognitive overload. Unifying continuous, streaming feature engineering with sub-100ms multi-class inference introduces severe latency bottlenecks and model miscalibration errors.
• Methodology & Architecture: This project evaluates a decentralized, serverless machine learning framework. It leverages AWS Kinesis for real-time interaction stream ingestion, executes distributed feature extraction via Apache Spark (AWS Glue), and serves continuous, live predictions using specialized XGBoost models optimized via temperature scaling to accurately map personal engagement vectors ($P(\text{click} \mid \text{send at hour } H)$).
• Key Benchmarks: Target validation focuses on maintaining strict systemic boundaries ($p99 \le 120\text{ms}$) within a specialized GraalVM-compiled Java runtime while minimizing Expected Calibration Error ($\text{ECE} \le 0.05$).

• Core Languages: Java (21+, GraalVM / Native Image Compilation), Python, TypeScript, SQL
• Cloud & Distributed Computing: AWS CDK, Lambda Architecture, Step Functions orchestration, EventBridge
• Data Ingestion & ML Infrastructure: Apache Spark, AWS Glue ETL, Amazon Kinesis Streams, SageMaker Inference Pipelines, XGBoost Calibration Ecosystems