University of Georgia
ML/AI Researcher
June 2025 – December 2026
Athens, Georgia
- Top results across all five homes. Achieved the highest accuracy and weighted F1 on all five smart homes and the highest macro-F1 on four of five (93.2 Aruba, 86.6 Cairo, 84.1 Kyoto, 93.2 Orange4Home; tied on Milan), as measured against 14 published baselines under one shared protocol averaged over five seeds, by splitting each activity episode into four views and training one lightweight expert per view.
- Gains over the strongest published baseline. Raised macro-F1 over the best published baseline on each home by up to 8.7 points (+8.7 Cairo, +4.1 Kyoto, +2.7 Orange, +1.3 Aruba), as measured on identical stratified 60/20/20 splits, by weighting each expert per episode by how often it was right on similar validation episodes.
- Efficiency. Beat the strongest multimodal baseline (CARE) by 3.7 to 20.6 macro-F1 points at about one-sixth of its inference cost, as measured by 591 versus 3,631 MFLOPs and 8.2M versus 11.2M parameters, by replacing joint multimodal training with independently trained experts and a parameter-free fusion rule.
- No trained fusion network needed. Outperformed four jointly trained fusion designs (Concat+MLP, fusion-token transformer, one-stage SupCon, fine-tuned two-stage alignment) by 1.3 to 2.0 pooled macro-F1 points, and a frozen two-stage alignment by 7.3, by combining frozen experts with local competence weighting, which also lets any expert be swapped without retraining.
- Showed where the gain comes from. Demonstrated that the improvement comes from episode-specific weighting, as measured by a net +121 correct predictions (328 errors fixed versus 207 introduced) on the 26.5% of episodes where experts disagree, and 10.3 percentage points more weight on correct experts, by running leave-one-expert-out and per-episode weight diagnostics.
Python
PyTorch
Machine Learning
Multimodal Fusion
Sensor Intelligence
TypeScript
JavaScript
Vue 2/3
React
Next.js
Tailwind CSS
Node.js
Express
MongoDB
RESTful APIs
Git
Jira
Postman
Swagger
PostgreSQL
Redis
Redux
Docker
Kubernetes
AWS
CI/CD
Prometheus
Grafana
HTML
CSS
PHP
WordPress
