What we study
Open Research Lab
Focus
MadGAA-LAB studies how AI agents behave in high-stakes, human-facing settings — starting with healthcare — and how to measure that behavior reliably.
We publish our tools, scenarios, and results openly so that others can reproduce, challenge, and build on them.
Research Areas
Multi-Agent Clinical Simulation
Simulated patients, clinicians, and examiners that interact through dialogue, used to train and test medical AI agents before real-world use.
- patient persona design and adversarial scenarios
- OSCE-style structured clinical examinations
- real-time voice interaction
Related projects: OSCE-Project, OSCE Real-Time Voice
Evaluation & Benchmarks
Methods for scoring agent behavior on empathy, persuasion, safety, and task success, with public leaderboards for fair comparison.
- LLM-as-judge rubrics and their reliability
- standardized scenario sets
- open leaderboards
Related projects: OSCE-AgentBeats Leaderboard
Healthcare Data Agents
Agents that retrieve and reason over structured clinical records using interoperability standards.
- FHIR database querying
- retrieval-augmented clinical reasoning
Related projects: FHIR-Agent
Agent Safety & Red-Teaming
Testing how agents hold up against prompt injection, manipulation, and other adversarial pressure.
- attack and defense scenarios
- robustness evaluation
Related projects: AgentBeats Security Arena
Efficient Models
Making capable models small and fast enough to run on local and edge hardware.
- knowledge distillation and domain fine-tuning
- hardware-aware optimization and inference benchmarking
Related projects: Model Training & Distillation
Work with us
We welcome collaborators from clinical, academic, and engineering backgrounds:
- propose a joint study or evaluation
- contribute code, scenarios, or rubrics on GitHub
- host a student project or internship
Reach us at hello@madgaalab.org or on GitHub.