AI Engineer, DeepAuto.ai

Agentic AI  ·  Generative Models  ·  Computer Vision

Email  ·  GitHub  ·  Google Scholar  ·  Blog

I focus on agentic AI and computer vision tasks with practical real-world applications: developing agentic AI systems that interpret piping and instrumentation diagrams (P&IDs) and transform them into structured engineering data, designing adversarial perturbations to disrupt deepfake synthesis, and building real-time human posture feedback systems, such as powerlifting form correction.

Engineering Diagram Digitization. (Ongoing research; findings have not yet undergone peer review.) P&IDs encode process topology through equipment, instruments, and their interconnections, but PDF representations often lack explicit machine-readable structure. I work on reconstructing this topology as an enriched knowledge graph, tracing pipe runs through reducers, T-junctions, and off-page connectors while tracking line continuity and identifying line-number changes. The goal is to automatically generate structured Line Listings and Instrument Listings from raw P&ID PDFs, including FROM/TO connectivity, line identifiers, operating pressure, and other engineering attributes.

Bilateral Asymmetry in Medical Imaging. (Ongoing research; findings have not yet undergone peer review.) Recent models that consider information from paired organs can improve overall predictive performance, but this benefit may not extend to patients with substantial left–right asymmetry. I investigate whether and when bilateral fusion improves population-level performance at the expense of predictions for asymmetric cases, and whether asymmetry-aware approaches can mitigate this trade-off without sacrificing overall accuracy.

Adversarial Protection against Deepfakes. I proposed an adversarial protection method that disrupts the unauthorized use of facial images in diffusion-based deepfake generation by applying perturbations to the original images before synthesis, rather than detecting manipulated content afterward. My latest method, AEGIS, is presented in a manuscript currently under review at Pattern Recognition. Earlier research in this direction received a Paper Award at the 2024 IEIE Autumn Annual Conference.

Low-light Object Detection. I proposed an approach to improve object detection in nighttime industrial environments, where poor illumination limits detection reliability. The approach combines diffusion-based data augmentation during training with low-light image enhancement at inference time, without modifying the underlying detector architecture.

Real-time Human Pose Estimation and Posture Correction. I proposed a real-time posture correction system using human pose estimation to assess exercise form and provide corrective feedback during the three major powerlifting movements: squat, bench press, and deadlift. Unlike post-workout video analysis, the system provides immediate feedback during exercise. This work was published in IEEE Access and has been cited 33 times.

Low-Light Object Detection — framework for object detection under the low-light and nighttime conditions of smart-yard environments; the system behind the IEIE journal paper.

Anti-StableDiffusion — adversarial method that disrupts deepfake generation through the Stable Diffusion img2img pipeline; the system behind the IPIU 2025 paper.

AI Exercise Pose Feedback — real-time powerlifting form correction with YOLOv5 and MediaPipe; the system behind the IEEE Access paper.

Paper Implementations — models from papers, reimplemented in PyTorch.

Study notes, mathematical foundations, and hands-on implementations of core AI methods.

Python  PyTorch  NumPy  pandas  scikit-learn  CUDA  Linux  Docker  AWS