My research develops Pragmatic AI: methods that stay reliable, controllable, and useful in the messy real world. It spans three directions.

1. Robust AI
AI that remains reliable under noisy, incomplete, distorted, or difficult real-world inputs.
Applications: Healthcare, Defense, Media, Transportation, Smart Cities, Software, Recruiting, Customer Service.

1. LLMs that understand noisy patient narratives
2. LLMs that ask useful follow-up questions
3. LLMs that interpret critical voice communications
4. Neural networks that recognize license plates from extreme angles
5. Neural networks that detect anomalies in cyber-physical vehicle systems
6. ML models that detect tunnels from weak seismic signals

2. Specification-Driven Generative AI
AI that generates outputs according to strict specifications, constraints, and target properties.
Applications: Material Science, Healthcare, Education, Media, Defense.

1. Diffusion models that design new molecules with desired properties
2. LLMs that generate text in a precise style
3. LLMs that generate text with controlled multi-aspect sentiment
4. LLMs that simulate students with specified knowledge gaps

3. Multi-Agent AI without Coordination
AI agents that solve shared tasks without prior coordination, direct communication, or centralized control.
Applications: Drones, Defense, ML Systems.

1. Drones that collaboratively allocate themselves across a pool of targets
2. LLMs that evaluate and rank each other’s outputs