Armin Zirak.

Architected and own a petabyte-scale data-observability framework for the catalog platform. It runs parity checks against external retailer data, classifies disparities, and routes disputes through a human review workflow that writes decisions back to production. Led a Snowflake cost-reduction initiative end to end that cut spend by $40,000+/month, plus a snapshot-modeling framework that saves another ~$400/day. Also delivered the Source-of-Truth data models the marketing and ML teams run on, which meant untangling business logic spread across many upstream systems and coordinating across those teams.

Re-architected the accounting and tax-reporting data marts — untangled tightly-coupled legacy logic into smaller, well-scoped pipelines and took major-impact incidents in that domain from about three a week to near zero. Designed the integration architecture for multi-currency payment-settlement and tax-file publishing (SFTP, S3, AWS Secrets Manager) to external banks and jurisdictions. Ran a lift-and-shift migration of ~120 tables from PostgreSQL to Snowflake with a view-swap cutover and a tested rollback path, and moved batch updates to Change Data Capture for granular change data — about $50K/year in savings.

Worked on a re-architecture of the partner reporting pipeline: a legacy batch pipeline replaced with Spark on AWS Glue, then a TypeScript + AWS SQS streaming path added alongside it so fresh attributes flow in low-latency while bulk analysis stays on batch. Contributed to the Redshift to sharded Aurora PostgreSQL migration and to query and sharding-key tuning — the effort cut application response times by ~80%, held dashboard latency sub-100ms, and lowered database and pipeline cost by ~40%. Also worked on the partner-facing and customer reporting APIs in Node.

M.Sc. research on language-model domain adaptation for automated program repair. Built distributed GPU training pipelines with Python, PyTorch, and Hugging Face for Transformer models (T5, BERT) and worked with the OpenAI APIs. First-author paper in ACM TOSEM, presented at ICSE 2024. Returned to the lab in early 2023 for a focused stint on OpenAI GPT fine-tuning for the same problem.

Built and evaluated machine-learning models for real-time stock-market price prediction, in partnership with the University of Calgary. Used Python and MLflow for the model lifecycle and experiment tracking.

Architected an end-to-end recommendation platform serving 20M+ unique users at 500+ RPS. Built low-latency backend microservices in Python (Django) with RabbitMQ, designed Redis caching layers, and optimized PostgreSQL queries to hold sub-100ms responses as traffic scaled. Built Spark pipelines and implemented the recommendation algorithms (collaborative filtering, matrix factorization, content embeddings); the recommendation feed reached ~2% CTR versus sub-1% on the platform's ads, validated through A/B testing.