LLMs · Security Research · Computer Vision
VSCode extension for on-demand C/C++ vulnerability detection. TypeScript extension + FastAPI backend, tree-sitter static parsing, in-house fine-tuned LLM over SSH tunnel. CodeLens, Diagnostics, and Hover UI. Presented at CyberQ 2025.
Unified CVE dataset with full metadata, commit context, and exploit signals. 2-stage tree-sitter pipeline across 120+ C repos for 0-day hunting. CPG-based graph datasets via Joern — CFG/DDG/CDG — with a per-node scoring algorithm.
Multi-judge evaluation system comparing proprietary and frontier models on correctness and completeness. BFCL benchmark coverage, structured-output and tool-calling tasks. Prompt-optimization pipeline using LLMs to discover optimal prompts per dataset.
Learning-augmented Artificial Potential Field scheme for narrow-corridor robot traversal. LoG edge detection, corridor-width estimation via chord length and law of cosines, dynamic repulsive field reduction. Deployed on Rosmaster R2 (Jetson Nano).
Image-based inspection pipeline benchmarking MobileNetV2, EfficientNetB0, ResNet50, ViT, and EfficientNet-L2 (SAM/ASAM). Led dataset curation — duplicate removal, relabeling, rebalancing. Showed data quality dominates architecture choice.
Comparative benchmark of CNN/transformer variants under unified preprocessing. Evidence that data curation matters more than model size for PV inspection tasks.
Dynamic repulsive-field adjustment from LiDAR-derived corridor-width estimates enabling safe traversal through physical bottlenecks. Modular design for wheeled and legged robots.
Open to research collaborations, interesting engineering problems, and roles at the intersection of LLMs and security.