Ankit Thepe

Key Skills

Software

CVAT

Label Studio

LabelImg

Top Subject Matter

Security/Surveillance (ATM anomaly detection)

Natural Language Processing

Travel/Tourism AI

Top Data Types

Video

Text

Image

Top Task Types

Bounding Box

Classification

Freelancer Overview

ATM Anomaly Dataset Curation - Data Annotator. Brings 2+ years of professional experience across complex professional workflows, research, and quality-focused execution. Core strengths include CVAT, Label Studio, and Internal. Education includes Bachelor of Technology, Madhav Institute of Technology & Science (2022) and Higher Secondary Certificate, World Way International School (2022). AI-training focus includes data types such as Video, Text, and Image and labeling workflows including Bounding Box, Classification, and Evaluation.

Labeling Experience

Software Developer Intern (AI focus): Human-in-the-loop Evaluator/Annotator

Text

I acted as a human evaluator of AI-generated travel itineraries, assessing their accuracy, tone, and logical flow. My feedback played a vital role in refining recommendation engines by providing reinforcement learning from human feedback. The work ensured robust personalization and improved user satisfaction. • Evaluated and rated AI-generated content • Provided structured feedback for RLHF processes • Focused evaluations on travel and itinerary domains • Participated in dataset curation for AI optimization

I acted as a human evaluator of AI-generated travel itineraries, assessing their accuracy, tone, and logical flow. My feedback played a vital role in refining recommendation engines by providing reinforcement learning from human feedback. The work ensured robust personalization and improved user satisfaction. • Evaluated and rated AI-generated content • Provided structured feedback for RLHF processes • Focused evaluations on travel and itinerary domains • Participated in dataset curation for AI optimization

2026 - Present

AI/ML Trainee - Dataset Labeler/Annotator

Label StudioTextClassification

I prepared, cleaned, and labeled datasets for NLP pipelines, focusing on text classification and related annotation tasks. The work included rigorous quality assurance to identify misclassifications and edge cases. These efforts contributed to the accuracy of downstream AI models. • Cleaned and labeled diverse text datasets • Applied annotation guidelines to ensure consistency • Performed manual review to identify and correct errors • Contributed to overall improvement of NLP model performance

I prepared, cleaned, and labeled datasets for NLP pipelines, focusing on text classification and related annotation tasks. The work included rigorous quality assurance to identify misclassifications and edge cases. These efforts contributed to the accuracy of downstream AI models. • Cleaned and labeled diverse text datasets • Applied annotation guidelines to ensure consistency • Performed manual review to identify and correct errors • Contributed to overall improvement of NLP model performance

2024 - 2025

MedCaption-LSTM Multimodal Alignment: Data Annotator

LabelimgImageClassification

I created paired text-image ground truth data by standardizing descriptive medical captions for use in training multimodal models. My work transformed unstructured medical data into structured datasets ready for model training. This process was essential for aligning image and text modalities for LSTM-based systems. • Paired images with standardized captions • Structured data for multimodal training scenarios • Ensured precise alignment of visual and textual data • Addressed challenges of unstructured medical imagery

I created paired text-image ground truth data by standardizing descriptive medical captions for use in training multimodal models. My work transformed unstructured medical data into structured datasets ready for model training. This process was essential for aligning image and text modalities for LSTM-based systems. • Paired images with standardized captions • Structured data for multimodal training scenarios • Ensured precise alignment of visual and textual data • Addressed challenges of unstructured medical imagery

Not specified

LLM Output Evaluation System (RAG): Output Rater/Annotator

Text

I validated semantic search contexts and rated generated answers from a large language model (LLM) for accuracy and grounding. This process helped improve model performance and minimize hallucinations. The work involved cross-verifying answers against source documents. • Rated LLM-generated outputs for quality • Cross-checked answers with original documents • Suggested prompt refinements for factual accuracy • Focused evaluations on factuality and relevance

I validated semantic search contexts and rated generated answers from a large language model (LLM) for accuracy and grounding. This process helped improve model performance and minimize hallucinations. The work involved cross-verifying answers against source documents. • Rated LLM-generated outputs for quality • Cross-checked answers with original documents • Suggested prompt refinements for factual accuracy • Focused evaluations on factuality and relevance

Not specified

ATM Anomaly Dataset Curation - Data Annotator

CVATVideoBounding Box

I manually labeled thousands of video frames with bounding boxes for object tracking and temporal action tagging tasks. This work supported the creation of training data for action recognition models such as SlowFast and X3D. Rigorous adherence to inter-annotator guidelines ensured consistent quality in challenging CCTV footage. • Labeled video data under variable lighting and occlusion conditions • Collaborated on dataset curation using agreed annotation standards • Supported development of action recognition algorithms • Verified and addressed challenging frames for quality control

I manually labeled thousands of video frames with bounding boxes for object tracking and temporal action tagging tasks. This work supported the creation of training data for action recognition models such as SlowFast and X3D. Rigorous adherence to inter-annotator guidelines ensured consistent quality in challenging CCTV footage. • Labeled video data under variable lighting and occlusion conditions • Collaborated on dataset curation using agreed annotation standards • Supported development of action recognition algorithms • Verified and addressed challenging frames for quality control

Not specified

Education

Work History