Research lines spanning computer vision, sustainable chemistry, livestock health, precision agriculture, and wildlife conservation.
ActiveNon-invasive detection of subacute ruminal acidosis via optical gas and thermal imaging
A multi-paper effort segmenting CO₂, methane, and multi-gas plumes from optical gas imaging — combining lightweight ViT backbones, attention-based fusion, and thermal recognition for non-invasive livestock health monitoring.
ActiveAutomated wildlife census and habitat analysis across camera-trap, thermal, and aerial imagery
Perception pipelines spanning ground-level RGB, thermal infrared, and drone-based orthoimagery to support large-scale ecological surveys and wildlife population studies.
ActiveMultimodal deep learning for accelerated discovery of sustainable solvent systems for biopolymer and plastic valorization
Vision-centric deep learning frameworks that fuse 3D imaging cues with cross-modal signals to forecast solvent–substrate interactions, enabling faster screening of sustainable solvent candidates.
Multi-stage weed detection, segmentation, and growth analysis with hierarchical transformers
Hierarchical and reparameterizable transformers, multi-task heads, and SAM-2 priors for real-time weed perception across diverse agricultural environments.
Lightweight transformer architectures for industrial detection and quality assessment
Lightweight transformer designs for real-time multi-class weapon segmentation, and dual-stream adversarial fusion for non-destructive frying-oil oxidation assessment.
CNN-based detection frameworks for cannabis seed identification and variant classification
Early-stage applied work on CNN detection pipelines for agricultural seed identification using RetinaNet and Faster R-CNN.
MESA — Comparative benchmarking of fine-tuned biomedical language models
Comprehensive 16-metric evaluation of BioGPT-Large, BioMedLM, and DialoGPT-Small fine-tuned with LoRA on 20K medical samples. Quantifies the trade-offs between clinical accuracy, patient-facing safety, and resource efficiency to guide model selection for healthcare deployment.
Joint MRI classification and segmentation with only 1.38M parameters
A compact multi-task network that performs simultaneous brain tumor classification and lesion segmentation on MRI scans. Combines deep supervision with a custom joint classification–segmentation loss, achieving 99.07% classification accuracy at a fraction of the parameter footprint of ViT-scale baselines (~1.5% the size).
Natural-language test automation for Android apps via vision-grounded LLM agents
An automated QA framework for the Obsidian mobile app built on Google's Agent Development Kit and Gemini Vision. A Supervisor–Planner–Executor agent hierarchy interprets plain-English test cases, drives the device through ADB, reasons over live screenshots, and produces structured pass/fail reports — no hand-written test scripts required.
Interactive segmentation-mask annotation tool for optical gas imaging datasets
A lightweight desktop utility for drawing and saving gas-region masks on images. Users trace boundaries with the mouse, fill enclosed regions, and the tool automatically generates paired binary masks and color overlays — purpose-built to bootstrap segmentation datasets for downstream training pipelines.