University of Cincinnati (Research)
AI/ML Research Assistant - Android Bug Reproduction
February 2026 - Present
- Designed a two-stage memory-augmented LLM method (Gemini 2.5 Pro video-to-memory extraction, then memory-conditioned device automation) that avoids repeated per-scene VLM inference
- Evaluated against ViBR (GPT-4o + GroundingDINO + CLIP) across 53 successful runs on 20+ real-world Android apps using a provider-agnostic LLM abstraction (Gemini, GPT-4o, Llama, Qwen, MiniCPM)
- Demonstrated 25.7x latency reduction (39.4s vs 1012.1s avg per run) over exhaustive scene-by-scene VLM replay
PythonGoogle Gemini 2.5 ProVertex AIADBuiautomator2scikit-image