Anugrha B.

Key Skills

Software

Roboflow

Top Subject Matter

YouTube thumbnail clickability prediction and image quality/active learning labeling

Computer vision automation with OCR-based data extraction

Predictive modeling for academic performance using tabular features

Top Data Types

Image

Top Task Types

Fine-tuning

Data Collection

Freelancer Overview

YouTube Thumbnail Clickability Predictor (In Progress). Core strengths include Hugging Face Spaces, OpenCV, and Jupyter Notebook. Education includes Bachelor of Computer Applications, Siliguri Institute of Technology (2030). AI-training focus includes data types such as Image, Computer Code, and Programming and labeling workflows including Fine-tuning, Data Collection, and Evaluation.

Labeling Experience

Behavioral Analytics & Academic Performance Predictor

Developed a model to predict academic performance (0-100) from behavioral and lifestyle features using a tabular neural network trained on an engineered feature set. Validated generalization with cross-validation and assessed predictive accuracy through error and goodness-of-fit metrics. Performed permutation feature importance and residual diagnostics to identify which features most strongly influenced predicted grades. • Trained and evaluated a tabular neural network regressor for grade prediction. • Conducted 5-fold cross-validation and computed RMSE on holdout splits. • Produced interpretability outputs via permutation feature importance and residual checks. • Deployed an interactive inference interface for instant predictions.

2025 - Present

Android Emulator Vision Automation System

ImageData Collection

Created an emulator-based computer vision automation system that detects in-screen resources and uses OCR to extract counter values during runtime, producing structured observations for downstream automation. The system operates across virtualized Android emulator states and uses YOLO object detection combined with OCR parsing to collect labeled-like visual readings. Implemented robust retry logic and defensive OCR fallbacks to ensure consistent data extraction under noisy, uncontrolled conditions. • Used YOLO detections to locate resource regions (e.g., Gold/Elixir/Dark Elixir counters). • Applied EasyOCR to parse numeric text for structured state readings. • Added producer-consumer GUI architecture to support continuous extraction. • Implemented multi-attempt retries and interrupt handling for reliability.

2025 - Present

YouTube Thumbnail Clickability Predictor (In Progress)

ImageFine-tuning

Built a genre-aware pipeline for generating labeled data from YouTube thumbnails using intersection-based image quality model outputs, then normalized labels per genre to reduce channel-size and topic bias. Fine-tuned a ResNet34 clickability classifier on a genre-balanced labeled set and added explicit handling for portrait-format out-of-distribution cases. Developed a human-in-the-loop active learning workflow using pairwise thumbnail comparisons and a Bradley-Terry ranking model to iteratively produce higher-quality retraining labels. • Labeled/produced training targets for thumbnail clickability via NIMA/MUSIQ/TOPIQ outputs. • Generated retraining data through active learning pairwise comparisons and ranking. • Performed dataset balancing across five thumbnail categories (200 thumbnails). • Deployed the pipeline with continuous data collection via confidence scores and comparison UI.

2025 - Present

Education

Work History