Production-grade AI systems spanning medical imaging, autonomous vehicles, satellite analysis, and real-time edge deployment.
↗
● Medical AI
Automated Medical Image Segmentation
U-Net system for brain tumor segmentation on BraTS 2021. Active learning pipeline reducing annotation time by 35%. Deployed across 5 hospitals via MONAI framework.
0.92 DSC Score
−30% Workload
5 Hospitals
U-NetMONAIActive LearningBraTS 2021
↗
● Medical AI
Tooth Detection, Segmentation & Numbering
Intelligent Dental Imaging System for Endodontic Surgery Prediction. Cascade R-CNN on OPG X-rays with contour analysis for overlapping tooth disambiguation.
96% Accuracy
−15 min/patient
+27% Numbering
Cascade R-CNNDental AIOPG X-raysCAD Integration
↗
● Medical AI
Medical Image Enhancement & Super-Resolution
MRI denoising with SNR improvement from 12 dB to 18 dB on the IXI dataset. CT contrast enhancement via CLAHE for liver lesion detection. Integrated into PACS.
12→18 dB SNR
−22% Prep Time
PACS Ready
Few-ShotCLAHE3D SegmentationIXI Dataset
↗
● Medical AI
Retinal Vessel Segmentation & Lesion Grading
Transformer-UNet hybrid for automated diabetic retinopathy grading on DRIVE & STARE datasets. Multi-scale feature fusion for accurate vessel tree extraction and microaneurysm detection.
0.94 AUC
97.2% Sensitivity
ViT-UNetDRIVE DatasetRetina AI
↗
● Computer Vision
3D Human Body Pose Estimation
Transformer-based pose prediction with robust occlusion handling. Real-time deployment via TensorRT with under 30ms latency on Human3.6M benchmark.
42mm MPJPE
<30ms Latency
−4% vs Baseline
TransformerTensorRTHuman3.6MReal-Time
↗
● Computer Vision
ViUNet: Vision Transformer UNet for Segmentation
Novel architecture merging U-Net's localization strengths with ViT's global context modeling. Multi-scale attention mechanisms for robust feature representation across diverse imaging modalities.
Multi-Scale Attn
Global Context
ViTU-NetAttentionCustom Architecture
↗
● Computer Vision
Multi-Modal Detection: Pose, Face & Hand
MediaPipe/OpenCV unified system for simultaneous pose, facial landmark, and hand gesture tracking. Low-light face detection optimized via histogram equalization. Deployed in AR applications.
Real-Time
3-in-1 Pipeline
MediaPipeOpenCVARLandmark Detection
↗
● Computer Vision
Gym Workout Recognition via ST-GCN
Spatiotemporal Graph Convolutional Network for real-time gym exercise classification. ONNX deployment enabling efficient inference on fitness applications with 88% accuracy on UCF101.
88% Accuracy
UCF101
ONNX Deployed
ST-GCNONNXAction Recognition
↗
● Computer Vision
Image Restoration: Real-ESRGAN & GFPGAN
Historical image restoration pipeline using Real-ESRGAN. Fine-tuned GFPGAN for high-fidelity facial reconstruction. HAT super-resolution model deployed in a production mobile application.
4× Super-Res
Mobile App
Real-ESRGANGFPGANHATMobile Deploy
↗
● Computer Vision
Few-Shot Fashion Recommendation System
Hybrid few-shot learning approach for visual fashion item retrieval and recommendation. Siamese networks combined with transformer embeddings for outfit compatibility scoring.
Few-Shot
Siamese Network
Few-ShotSiamese NetTransformers
↗
● Detection
Thermal Object Detection: YOLO–Transformer Hybrid
Real-time colormap-invariant thermal detection framework. Hybrid YOLO–Transformer architecture achieving robust detection across varied thermal colormap configurations in industrial environments.
Colormap-Invariant
Real-Time
Thermal ImagingYOLOTransformerIndustrial AI
↗
● Detection
License Plate Recognition with YOLOv7
End-to-end plate detection and OCR recognition pipeline using YOLOv7. Optimized for varying lighting, angles, and international plate formats with high-speed inference.
YOLOv7
Multi-Format
YOLOv7OCRLPRReal-Time
↗
● Detection
Multi-Object Tracking with DeepSORT & ByteTrack
Robust multi-object tracking pipeline combining YOLOv8 detection with DeepSORT and ByteTrack. Real-time speed estimation and trajectory analysis for surveillance and traffic monitoring.
MOTA 78.4%
30+ FPS
DeepSORTByteTrackYOLOv8Tracking
↗
● Detection
Industrial Defect Detection System
Automated quality control pipeline for manufacturing defect detection. Anomaly detection via PatchCore and supervised YOLOv8 for surface scratch, dent, and contamination classification.
99.1% Precision
Zero False-Pass
PatchCoreAnomaly DetectionManufacturing
↗
● Satellite / Remote Sensing
Scalable High-Res Object Detection via Sentinel
Rare-object detection and matching framework for large Sentinel satellite tiles. Sliding-window + attention-guided detection with geospatial coordinate output and GeoJSON export.
Large Tile
GeoJSON Output
SentinelGDALRasterioTorchGeo
↗
● Satellite / Remote Sensing
Road Segmentation via Maxar Open Data & SAM2
Zero-shot and few-shot road network extraction from high-resolution Maxar imagery using Segment Anything Model 2. Automated pipeline for urban planning and infrastructure mapping.
SAM2
Zero-Shot
SAM2MaxarGeoPandasRoad Network
↗
● Satellite / Remote Sensing
Fire & Smoke Detection — Industrial & Wildfire
Dual-domain detection system for fire and smoke in both industrial environments and wildfire satellite imagery. Temporal analysis for early-warning systems with drone integration capability.
Early Warning
Multi-Domain
Fire DetectionTemporal CVDrone Integration
↗
● Satellite / Remote Sensing
Change Detection in Multitemporal Satellite Imagery
Siamese network architecture for detecting land-use and structural changes between multitemporal satellite image pairs. Deployed for flood damage assessment and urban expansion monitoring.
Siamese Net
Flood Assessment
Change DetectionSiameseMultitemporal
↗
● Edge AI
YOLOv11-EdgeSuite: Universal Mobile Vision Toolkit
Android mobile vision toolkit built in Kotlin using YOLOv11 for detection and segmentation on smartphones. Real-time inference via TensorFlow Lite with camera API integration.
Android
Real-Time
TF Lite
YOLOv11KotlinTF LiteAndroid
↗
● Edge AI
Edge/Corner Detection on Raspberry Pi 4
GPU-accelerated Harris Corner and Sobel edge detection for real-time video processing on Raspberry Pi 4. Adaptive thresholding for robust low-light performance.
GPU Accelerated
Real-Time Video
Harris CornerSobelRaspberry Pi 4OpenCV
↗
● Edge AI
Autonomous Driving Perception — CARLA + Openpilot
Full AV perception stack integrating object detection, lane segmentation, and depth estimation within the CARLA simulator. Openpilot-compatible policy evaluation for urban and highway scenarios.
CARLA Sim
Full Stack
CARLAOpenpilotDepth Est.Lane Seg
↗
● Edge AI
Sign Language Recognition — Real-Time ASL
MediaPipe Holistic + LSTM temporal classifier for real-time American Sign Language recognition. Deployed on embedded hardware with under 20ms latency and 94% word-level accuracy.
94% Word Acc.
<20ms Latency
MediaPipeLSTMASLEdge Deploy
