Software Engineering Intern
GDIT · General Dynamics Information Technology, Intelligence & Homeland Security
- Built a machine-learning classifier that detects prompt-injection and jailbreak attacks, reaching 96% recall on 543 unseen attacks.
- Found and repaired a 46% label-corruption problem in the training data through iterative retraining.
- Added uncertainty-based guardrails that flag unfamiliar attacks, and wrote a proposal that was adopted for the next generation of internal guardrail systems.
- Shipped a Flask web app demonstrating the full detection pipeline to stakeholders.