Samuel Kapusta

ML Software Engineer II · San Diego, CA

• ›Trained, labeled, and benchmarked sparse autoencoder architectures that decompose customer behavior embeddings into interpretable audience segments. Drove cross-org adoption of the winning model by Amazon Ads as the backbone of their enterprise advertiser product.
• ›Led a team of 4 building the production pipeline: Step Functions orchestrating daily SageMaker inference, monitoring and alarming on delivery, and a Redshift store holding activations across 4,000 features for all 330M customers.
• ›Built the data layer behind Amazon's customer foundation model — a real-time stream unifying 50M+ daily behavioral events into per-customer sequences, vending embeddings to downstream payment risk, ads, and package consolidation systems.

Data Science Intern · Basking Ridge, NJ

• ›Built Tableau dashboards over gateway metrics and geomapped session data to compare nationwide traffic, and automated Python jobs loading gateway data hourly into PostgreSQL.

Software Engineering Intern · San Diego, CA

• ›Built a desktop application used by university administrators to group and query secure student data.

An LLM chess coach that explains why a player loses rather than what the engine would play. Maia gives a human move-prediction baseline at the player's rating, Stockfish gives the objective evaluation, and a 183-tag rule engine names the mistake mechanism — Missed Zwischenzug, Hung Piece — before the model writes a word.

Python · PyTorch · React · Lambda · Fargate · SQS · DynamoDB · CDK