Project details

Full pipeline breakdown: data, model, evaluation, and deployment.

Three detection projects covering different domains and constraints. Each section documents the dataset, training setup, key challenges, and what I'd improve with more time.

80%10-class driver action accuracy

86.8%Fish detection mAP50-95

91.8%Meter ROI mAP50-95

VisionEdge / Edge AI

01 / Driver Safety

VisionEdge Driver Safety

In-cabin driver monitoring: detecting distraction, phone use, drowsiness, and unsafe postures from a single camera feed in real time.

Problem

Fleet operators need automated distraction alerts but can't afford cloud latency — detection must run on-device under 100ms per frame.

Solution

Trained YOLOv8m on 8,000+ augmented frames (10 action classes). Preprocessing: histogram equalization, random crop, mosaic. Exported to ONNX for Jetson Nano inference at 15 FPS.

Result

80% accuracy across 10 classes. Main challenge: low-light and motion blur in real cabin footage — addressed with aggressive augmentation and contrast normalization.

PythonYOLOv8mOpenCVONNXJetson NanoColab T4

AquaSight aquaculture intelligence prototype

AquaSight / Detection + Counting

02 / Aquaculture

AquaSight Fish Counter

Underwater object detection for fish counting in dense, occluded, and variable-lighting conditions. Built a full operator interface around the model.

Problem

Manual fish counting is slow (2+ hours per pond) and inaccurate under turbid water, overlapping bodies, and inconsistent lighting.

Solution

Annotated 2,000 images (LabelImg), trained YOLOv11m for 120 epochs on Colab A100 (batch 16, img 640). Built PyQt5 dashboard with live count, confidence filter, and CSV export.

Result

86.8% mAP50-95. Key challenge: heavy occlusion in dense schools — mitigated with mosaic augmentation and NMS tuning (IoU 0.45).

YOLOv11mPyQt5LabelImgPythonColab A100OpenCV

Meter ROI

mAP 91.8%

DocuFlow / Detection + OCR

03 / Document AI

DocuFlow Meter OCR

Two-stage pipeline: YOLO detects meter/field regions, then PaddleOCR extracts digits from cropped ROIs. Handles skew, blur, and uneven lighting.

Problem

Utility workers photograph meters in poor conditions (glare, angle, dirt). Manual transcription has 8-12% error rate and takes 3x longer than automated reading.

Solution

Collected 2,200 document samples. Trained YOLOv8 for ROI detection (meter display, serial number, date fields). Applied perspective correction + adaptive thresholding before PaddleOCR extraction.

Result

91.8% mAP50-95 on ROI detection. Challenge: digit confusion on damaged/faded displays — improved with synthetic data augmentation (blur, noise, fade).

YOLOv8PaddleOCROpenCVPythonRoboflowColab T4

What's next

I'm looking for a team where I can apply these skills on real-world data at scale.

Get in touch