Soham L.

Key Skills

Software

Other

Roboflow

LabelImg

CVAT

Top Subject Matter

Computer Vision

Data Annotation & Labeling

Maritime Security

Top Data Types

Image

Text

Top Task Types

Segmentation

Bounding Box

Classification

Object Detection

Fine-tuning

Cuboid

Evaluation/Rating

Freelancer Overview

Computer Vision & Data Annotation specialist with experience in underwater image enhancement, object detection (YOLO, U-Net), and Edge AI deployment (ESP32-CAM). Proficient in TensorFlow, Keras, PyTorch, Docker, and FastAPI. Handled image and text datasets with labeling workflows including bounding boxes, segmentation, classification, and evaluation.

Labeling Experience

AI-Generated Text Evaluation via RL Training Environment

OtherText

I built and launched an RL-based evaluation platform where AI agents assess and rate AI-generated text based on set rubrics. This involved designing reward systems, automating assessment criteria, and reproducibly benchmarking agent outputs. The experience required rigorous evaluation of AI outputs and structured feedback cycles using predefined metrics. • Created custom rubrics for scoring and evaluation of AI-generated text. • Integrated sentence-transformer similarity metrics for automated rubric assessments. • Ran RL training loops, providing continuous evaluation on agent improvements. • Deployed the evaluation pipeline as a live API for real-time testing and session tracking.

2026 - 2026

Underwater Image Enhancement and Object Detection for Maritime Security

ImageSegmentation

I developed an AI system to enhance and label underwater images for the detection of threats in low-visibility conditions. The work involved training and fine-tuning a U-Net model for image enhancement, as well as a YOLOv8 object detection model for identifying threat classes. Model performance was rigorously evaluated to ensure only high-quality labeled outputs were used in production. • Annotated and pre-processed underwater images for segmentation and object detection tasks. • Fine-tuned the YOLOv8 model to identify multiple threat categories from images. • Used leaderboard-based quality gates for automated evaluation and acceptance of labeled data. • Enabled real-time feedback on labeling quality through a live web app interface.

2026 - 2026

Robotic Arm – AI-Based Object Detection & Sorting

ImageObject Detection

Annotated 500+ real-world images for Edge AI deployment on ESP32-CAM — labeled 5 object classes using LabelImg with strict quality control, enabling 95% real-time detection accuracy on resource-constrained hardware Optimized annotations for Edge AI constraints — ensured tight bounding boxes, handled occluded objects, and maintained consistency across poor lighting and varying angles Applied data augmentation (rotation, HSV jitter, brightness, flip) to improve model robustness without additional manual labeling

2025 - 2025

Education

Work History