Built an orchard-scale apple detection study focused on visual reliability under occlusion, lighting variation, dense fruit clusters, and leaf-cluttered backgrounds. Associated paper accepted at ICICV 2026.

• Formulated single-class apple detection on AppleBBCH81 for real-world orchard imagery.
• Benchmarked one-stage, two-stage, and transformer-based detectors using COCO-style mAP, precision, recall, F1-score, and precision-recall analysis.
• Studied edge deployment feasibility for Raspberry Pi-class devices, focusing on models that remain useful outside clean benchmark settings.

Developed a silhouette-based gait recognition framework for identity matching when face cues are weak or unavailable. The work focuses on temporal modeling under view, clothing, carrying, noise, and long-sequence variation.

• Designed a CNN + Temporal KAN architecture with learnable temporal dynamics and gated long-term memory.
• Modeled local gait cycles and broader motion context for robust identity matching across surveillance conditions.
• Validated the framework on CASIA-B using gallery-probe matching across normal walking, bag-carrying, and clothing-change settings.

Worked on mammography ROI classification, focusing on benign-malignant lesion understanding under variation in texture, scale, appearance, and class balance.

• Built a hybrid EfficientNetV2-M + Vision Mamba pipeline for local lesion representation and efficient global context modeling.
• Evaluated the system on CBIS-DDSM lesion crops against CNN and Transformer baselines.
• Connected model design with practical diagnostic constraints, where subtle visual patterns and long-range context both matter.

Worked on applied AI systems for enterprise data access, internal document intelligence, forecasting, and workflow automation.

• Developed an agentic AI system for querying structured, unstructured, and operational data through natural language.
• Built prediction, NLP, and automation components for internal decision-support workflows.
• Contributed to production-oriented tools where reliable retrieval and usable outputs mattered more than demo performance.

Explored reward-guided adaptation of Stable Diffusion v1.5 for task-specific image generation, using reinforcement learning to steer visual quality beyond the base diffusion objective.

• Designed a parameter-efficient RL fine-tuning setup using LoRA layers in U-Net attention blocks.
• Used reward scoring and PPO-style policy updates to guide latent diffusion outputs.
• Analyzed the difference between moderate reward optimization and over-optimization in generated images.