Building data
systems that
stay resilient
when the real
world gets messy.

Global Infrastructure Services · Richmond, Virginia

• Established a pilot-ready Azure landing zone for the CDC NBS 7 public-health modernization initiative by provisioning AKS, HDInsight Kafka, SQL Server, Blob Storage, Azure Files, Container Registry, and supporting services through reusable Terraform configurations.
• Secured communication across the distributed environment by implementing virtual networks, private endpoints, DNS, NAT, firewall controls, Application Gateway, Key Vault, and container-image delivery between ACR and Kubernetes workloads.
• Enabled operational visibility by integrating Azure Monitor, Prometheus, and Grafana, then performed end-to-end testing across AKS deployments, Kafka publish-subscribe flows, SQL connectivity, storage access, ingress routing, outbound communication, and log ingestion.
• Advanced the environment from proof of concept to pilot readiness by resolving integration gaps and delivering reusable infrastructure modules, Helm configurations, network documentation, validation results, deployment runbooks, and readiness checklists.

Clarkson University · Potsdam, New York

• The department needed a reference-quality, production-style data platform demonstrating real-time and historical weather analytics, built from raw station telemetry with no tolerance for silent data loss.
• Tasked with designing and building the complete pipeline — ingestion, processing, storage, APIs, and monitoring — end to end as a solo data engineer.
• Engineered an event-driven platform using Python, Apache Kafka, PostgreSQL, Docker, and Kubernetes: built idempotent producers/consumers with offset commits gated on successful database writes, automated backfills with timestamp-based reconciliation across source, producer, consumer, and database states, REST APIs and a dashboard for live and historical trends, and Prometheus/Grafana observability across the stack.
• Delivered a fully containerized, self-healing platform that recovers only missing observations after any service or infrastructure interruption, with automated lifecycle workflows for deployment, backfill selection, and persistent storage across restarts.

Clarkson University · Potsdam, New York

• Investigated prolonged EMS on-scene time in stroke emergencies by analyzing the 2022 NEMSIS dataset of 51M+ nationwide activations; developed a preprocessing workflow that filtered ground-transport stroke cases, prevented target leakage, and produced a modeling cohort of 116,504 complete records.
• Addressed high dimensionality and non-normal feature distributions by applying Anderson-Darling testing, Spearman correlation analysis, near-zero-variance filtering, Chi-Square tests, and Kruskal-Wallis tests, reducing 32 candidate variables to 19 statistically relevant predictors.
• Resolved a 13.5% minority-class imbalance using ROSE synthetic sampling and an 80/20 stratified train-test split; trained and compared Logistic Regression, Decision Tree, and Random Forest models, with Random Forest achieving 70.14% accuracy, 93.65% recall, 70.25% precision, and a 0.802 F1 score.
• Evaluated model coefficients and feature importance to identify patient-handling time, total call duration, medication activity, age, response characteristics, and geographic region as key contributors to prolonged scene time, generating evidence-based findings relevant to prehospital stroke-care optimization.

Amazon · Hyderabad, India

• Prepared and labeled training data for a product listing approval system that triaged 4,000+ daily product submissions into approve, reject, or pending decisions — enabling automation that reduced the manual review backlog by 85% within the first deployment cycle.
• Identified that the deployed system was generating incorrect rejections at scale across 4 marketplace databases, flagging a critical gap between training data quality and real-world model performance that was causing a surge in seller appeals.
• Analyzed 300+ weekly seller appeal cases to surface recurring model failure modes and SOP gaps, working directly with sellers to help them get reinstated while building a structured evidence base for proposed fixes.
• Proposed targeted changes to both the model's decision logic and the appeals process, contributing to a 19-point accuracy improvement from 46% to 65% and reducing false-positive rejection rates from 54% to 35%.

Amazon · Hyderabad, India

• Audited samples of 30K+ weekly customer complaints classified by an automated keyword-based bucketing system, submitting true/false positive assessments that fed back into retraining and improving the classifier over time.
• Investigated 200+ seller, buyer, and fulfillment records weekly across 4 regional marketplace databases, tracing wrong item complaints to three root causes — inventory mismatches, seller fraud through misleading listings, and buyer abuse of the returns process.
• Developed evidence-backed enforcement recommendations including seller restrictions, bans, fraudulent buyer flags, and inventory audit triggers, and presented them to stakeholders for approval.
• Tracked complaint volumes over 6–8 weeks post-implementation to validate impact — reducing marketplace complaint rates by 45% by escalating 30+ fraudulent seller accounts and flagging 47 high-risk buyer profiles.