Projects

Here are some of the key projects I have led or made significant contributions to. My work focuses on applying advanced AI and Machine Learning techniques to solve complex industrial and research problems.

Agentic AI Routing System for Industrial Processes

As the lead on this project at Armada, I architected a scalable agentic AI routing system designed for complex industrial applications. The system leverages advanced ML and optimization to translate complex data into actionable, natural-language insights for operators.

Key Technologies: Generative AI, LLMs, Reinforcement Learning (GRPO), Self-correction Mechanisms, Python.
Impact & Achievements:
- The distilled models outperform large-scale proprietary systems by over 10%.
- The system is optimized for edge deployment, ensuring high performance in real-world industrial environments.
- Improved overall efficiency of complex industrial processes by a projected 4.5%.

Deep Neural Network for Flight-Level Traffic Forecasting

At American Airlines, I led the development of a deep neural network forecasting engine to predict flight-level traffic for the Yield Management team. This involved using a multimodal approach to capture complex trends and seasonality.

Key Technologies: Deep Learning, Multimodal Neural Networks, Time-Series Analysis, Python, PyTorch, AzureML.
Impact & Achievements:
- Reduced prediction mean square errors by 50% compared to previous models.
- Successfully deployed for A/B testing and potential replacement of legacy systems.
- Expanded to cover over 7,000 flights, providing accurate predictions for the next 330 days.

Multimodal Deep Learning for Robotic Control & QA

This work, stemming from my doctoral research and applied at Armada, focuses on integrating language and vision to create more intelligent systems.

Key Technologies: Multimodal Deep Learning, Reinforcement Learning, NLP, Computer Vision (CNNs), Python, PyTorch.
Impact & Achievements:
- Pioneered OPUS, an LLM framework to control robotic camera systems with natural language, achieving 20% higher task accuracy than large proprietary models.
- Developed a novel multimodal architecture that reduces mean squared error by 10-15% across diverse tasks.
- Introduced a Textual Question Answering architecture that employs on-demand visual grounding, achieving performance comparable to GPT-4 with a 400M parameter model.

Ontology Alignment and Clinical Data Processing

While at Truveta, I contributed to the mission of “Saving Lives with Data” by developing data-driven NLP approaches for clinical text.

Key Technologies: Large Language Models (LLMs), NLP, Ontology Alignment, Python.
Impact & Achievements:
- Pre-trained and fine-tuned LLMs on clinical data, significantly increasing downstream training and inference speeds (fourfold and twofold, respectively).
- Achieved state-of-the-art results in ontology alignment, with a 5% improvement in F1, Hit@1, and MRR.
- Developed Truveta Mapper, a novel framework for unsupervised ontology alignment.