Enterprise AI,
engineered for reality.

Building ML infrastructure and reproducible evaluation workflows for protein-language-model and multi-omics use cases.

• Build ML infrastructure and evaluation workflows for protein-language-model and multi-omics use cases within Amgen's generative-biology environment.
• Benchmark AMPLIFY's 120M- and 350M-parameter checkpoints on protein sequences up to its published 2,048-residue context.
• Track task metrics, model configurations, sequence-length test slices, and efficiency regressions in MLflow to make checkpoint comparisons reproducible.
• Analyze task-quality and compute trade-offs between the 120M and 350M checkpoints by sequence-length slice; the published AMPLIFY 350M comparison reports 43× fewer parameters and 24–29× higher inference throughput than ESM2-15B, depending on sequence length.
• Engineer Spark SQL and Delta Lake pipelines for multi-omics data managed through AWS HealthOmics and Databricks.
• Package selected model services for controlled deployment with Amazon Bedrock, Docker, and Kubernetes.

Contributed to retrieval, evaluation, computer-vision, and data workflows during an 11-month graduate internship.

• Contributed to retrieval and evaluation workflows for Ask AI during its 2025 rollout across Albertsons banner apps; Google reported that 85%+ of conversations began with open-ended or exploratory queries.
• Built labeled evaluation sets covering exploratory shopping intents, product-discovery requests, and ambiguous natural-language queries.
• Measured retrieval relevance, response grounding, and answer quality across intent and product-category slices before release.
• Prototyped computer-vision workflows for produce-quality inspection with distribution and engineering teams.
• Prepared BigQuery, Spark SQL, and Delta Lake datasets for merchandising and promotional analysis.
• Implemented request validation, PII controls, and rate limiting for customer-facing FastAPI services.

Built explainable credit-risk models and distributed transactional pipelines for an institutional debt marketplace.

• Developed and compared XGBoost and LightGBM credit-risk models using transaction and behavioral features.
• Added SHAP-based reason codes, feature-level explanations, and model documentation for governance review.
• Supported workflows for a platform that processed 9M+ transactions in the first nine months of FY24 and reported 40% year-over-year GTV growth; these are platform-level figures.
• Built Kafka and Snowflake pipelines for transaction normalization and behavioral feature generation.
• Automated repeatable training and evaluation workflows with Databricks, Spark SQL, MLflow, and Airflow.
• Packaged selected models behind Dockerized FastAPI services for integration with downstream applications.

Implemented data, analytics, and model-delivery workflows for enterprise client projects.

• Implemented Spark-based AWS ETL and reporting workflows that consolidated multi-region data for KPI reporting and predictive analytics.
• Prepared and transformed model-ready datasets for supply-chain forecasting workflows built with Scikit-learn.
• Supported fraud-detection experiments using Random Forest and XGBoost, including feature preparation and model evaluation.
• Tuned PostgreSQL queries and added Redis caching where repeated analytics requests were slowing dashboards.
• Supported containerized model delivery and CI/CD workflows for repeatable development and release processes.
• Built Tableau and Matplotlib reports for KPI, pipeline, and model-review discussions with client teams.

Developed physiological signal-processing, time-series modeling, and inference workflows for contactless patient monitoring.

• Built ballistocardiography signal-processing pipelines that converted under-mattress micro-vibration data into cardiac and respiratory features; by June 2021, the Government of India reported that the broader Dozee platform had served 30,000+ patients across 4,000+ beds in 35 districts.
• Developed reusable preprocessing and feature-extraction workflows with SciPy, Pandas, and NumPy for physiological time-series data.
• Trained and evaluated 1D-CNN and LSTM models for respiratory events and physiological anomalies.
• Compared model outputs with patient-level reference monitoring data and performed clinically relevant error analysis.
• Added temporal trend features used by early-warning workflows for longitudinal physiological monitoring.
• Packaged reusable feature extraction in Dockerized Flask services for low-latency inference.