Vivek K.

Key Skills

Software

No software listed

Top Subject Matter

Remote sensing / satellite imagery segmentation

Vision-language modeling for query-conditioned reasoning (VQA)

Deepfake detection and explanation generation

Top Data Types

Image

Top Task Types

Segmentation

RLHF

Fine-tuning

Freelancer Overview

Summer Research Intern — Natural language Interpretation of Satellite Imagery (ISRO/INTER IIT TECH MEET 14.0). Brings 2+ years of professional experience across complex professional workflows, research, and quality-focused execution. Core strengths include Hugging face and Pytorch. Education includes Bachelor of Technology, Indian Institute of Technology (BHU), Varanasi and Minor Program, Indian Institute of Technology (BHU), Varanasi. AI-training focus includes data types such as Image and labeling workflows including Segmentation, RLHF, and Fine-tuning.

Labeling Experience

Summer Research Intern — Natural language Interpretation of Satellite Imagery (ISRO/INTER IIT TECH MEET 14.0)

ImageSegmentationRLHF

Built and fine-tuned an image segmentation model for remote sensing use cases, using IoU preference loss to improve visual grounding quality. Adapted SAM3 to segment satellite imagery more accurately and consistently for downstream analysis tasks. • Fine-tuned SAM3 for remote sensing segmentation. • Used IOU preference loss to guide preference learning. • Performed query-conditioned reasoning with an image-language model pipeline. • Integrated models into an end-to-end satellite image understanding workflow.

2024 - 2024

INTER IIT TECH MEET 13.0 — Deepfake detection and explanation generation (model training and SFT-style instruction tuning)

ImageFine-tuning

Developed and trained an image-based deepfake detection model and produced artifact-related explanations. Created a large synthetic dataset using class-specific prompts to support robust instruction-driven fine-tuning and evaluation of detection and explanation performance. • Built a 2xEfficientNet + VAE classifier using reconstruction error channels. • Generated a 300K+ synthetic fake dataset with artifact explanation prompts. • Fine-tuned a vision-language instruction model for artifact identification and explanation. • Validated explanations via attention mapping.

2023 - 2023

Education

Work History