Where I've worked
Experience
Four years building and shipping machine learning systems in industry and research.
Research Assistant
University of Denver, Ritchie School of Engineering & Computer Science
- ▸Efficient Foundation Models: researching model compression, quantization, and hardware-aware optimization for on-device vision-language and multimodal inference. Building a benchmark dataset and efficient models for hardware e-waste detection and recovery in data centers, with CUDA-accelerated components.
- ▸Facial Expression Recognition: designing vision architectures trained with soft labels on AffectNet+ to address ambiguity between overlapping expressions and demographic bias.
- ▸Embodied Multiparty AI: designing multiparty perception for a social robot, fusing microphone-array direction into camera video to resolve who is speaking and who is addressed, with a geometry-grounded benchmark.
- ▸Speech Synthesis via Search: developing a black-box search framework that clones voices onto frozen TTS models for a social robot, raising speaker-similarity fitness from 84.95 to 96.61 in 2,000 evaluations. Now in production.
Data Scientist
BeanSprout AI
- ▸Built a cardiovascular risk prediction model using LightGBM and XGBoost, reaching AUC 0.89, and took it from data ingestion through hyperparameter tuning to deployment on Modal with MLflow tracking.
- ▸Implemented real-time speech interaction using Vertex AI speech-to-text and text-to-speech, and designed agentic workflows for secure document access and intent routing on an elder-care platform.
- ▸Engineered a RAG pipeline over patient documents with persistent chatbot memory and user accounts, removing repeated manual data lookups from each session.
Data Scientist
VHB (Vanasse Hangen Brustlin Inc.)
- ▸Analyzed terabytes of city-scale traffic data using PySpark, SQL, and Delta Live Tables on Databricks, and audited streaming feeds to find that over 70% of the data came from a single unreliable vendor.
- ▸Deployed an internal RAG chatbot with SQL execution, giving project managers natural-language access to employee hours, assignments, and reporting data that previously took days to pull manually.
- ▸Built a real-time object detection and tracking pipeline using YOLO and Azure Vision to analyze live intersection feeds and flag near-miss incidents, optimizing inference latency to keep up with streaming video.
- ▸Developed LSTM, Prophet, and SARIMA time-series models to forecast energy demand for NY Transit Authority, improving load-prediction accuracy to support peak-reduction strategies.
Data Scientist Intern
Kibur Medical Inc.
- ▸Built a bilateral filtering preprocessing pipeline for high-dimensional biomedical images, improving cell quantification accuracy by roughly 20%.
- ▸Integrated AWS S3 for scalable, real-time image retrieval, replacing local storage and supporting a continuous clinical imaging data workflow.
Research Assistant
Northeastern University
- ▸Trained and evaluated LSTM, Gaussian mixture, and regression models on time-series sensor data for COVID-19 detection, reaching 95% recall under noisy measurement conditions. Built training pipelines in PyTorch.
- ▸Applied detection theory, including MAP, Bayesian, Neyman-Pearson, and Minimax estimators, to separate true sensor response from background noise in low signal-to-noise biosensor measurements.
- ▸Published 4 peer-reviewed papers in IEEE Sensors Letters, Biosensors and Bioelectronics, IEEE CICC, and Journal of Applied Physics.