Ivan Koptiev

Description:
SERP AI is teaming up with Boxing Undefeated to create the first open, comprehensive boxing dataset focused on punch recognition and classification. We are labeling boxing match videos to power future ML models for real-time, open-source punch tracking and analytics for gyms, trainers, analysts, and fans.

• Project's GitHub page: Project
• LinkedIn page of a founder: Devin Schumacher
• LinkedIn page of a company: SERP AI
• Direct impact on the boxing community
• Experience with annotation tools, dataset creation, model training, and team/project management
• Contribution to a real world project

• Responsibilities: video annotation, label punches and boxers, quality control, preparing data for model training, managing annotation team.

Status: Active Start date: August 2025

Summary:
Built a text classification pipeline for news articles using bag-of-words and n-gram features. Compared LinearSVC, XGBoost, and AdaBoost models. LinearSVC achieved 86.9% accuracy, outperforming the others. Visualizations include class distribution and model comparison.

• Dataset: News articles (4 categories, balanced)
• Techniques: NLTK preprocessing, n-grams, model comparison
• Key learning: LinearSVC is highly effective for n-gram text classification

Summary:
Developed a deep convolutional neural network to classify images from the CIFAR-10 dataset. The model uses batch normalization, dropout, and real-time data augmentation to improve generalization. Achieved 82.8% accuracy and 83.8% validation accuracy after 50 epochs.

• Dataset: CIFAR-10 (60,000 images, 10 classes)
• Techniques: Batch normalization, dropout, data augmentation
• Key learning: Regularization and visualization are essential for deep learning success

Summary:
Built a K-Nearest Neighbors classifier for the scikit-learn wine dataset, systematically testing K values, distance metrics, and the effect of scaling. Achieved 98.15% accuracy with optimal parameters and demonstrated the dramatic impact of feature scaling. Visualizations include K value analysis, confusion matrix, scaling comparison, class distribution, and feature pairplot.

• Dataset: scikit-learn wine (178 samples, 13 features, 3 classes)
• Techniques: KNN, scaling, cross-validation, feature analysis
• Key learning: Scaling is critical for distance-based algorithms

Summary:
This project implements a Deep Q-Network (DQN) agent to solve the Acrobot-v1 control problem. The agent uses a three-layer neural network, experience replay, and target networks to achieve the solved threshold of -99.83 average reward. Key insights include the importance of balanced architecture, patient exploration, and training stability techniques like gradient clipping.

• Environment: Acrobot-v1 (double pendulum)
• Network: 6 → 64 → 64 → 32 → 3
• Best average reward: -99.83 (solved!)
• Key features: experience replay, target networks, gradient clipping

Summary:
Explores unsupervised learning with DBSCAN on the classic Iris dataset. The project features automated epsilon selection using the elbow method, comprehensive cluster visualizations, and a direct comparison to true species labels. Results include multi-dimensional plots, cluster statistics, and insights into noise handling and parameter tuning.

• Automated epsilon selection (elbow method)
• Clustered Iris data into 2 groups + noise
• Visual comparison to true species labels
• Learned about density-based clustering, parameter tuning, and noise handling

Summary:
Built a text classification pipeline for the 20 Newsgroups dataset using TF-IDF features and Multinomial Naive Bayes. The model achieved 91.2% accuracy and provided insights into class separability and feature importance. Visualizations include class distribution, confusion matrix, and top words per class.

• Dataset: 20 Newsgroups (4 categories)
• Techniques: TF-IDF, n-grams, feature importance analysis
• Key learning: Text preprocessing and feature engineering are critical for NLP

Summary:
This project demonstrates BERT for text classification and question answering, including fine-tuning on a real sentiment dataset. The model predicts sentiment for new texts and answers questions using transfer learning. Demo predictions and a bar plot of sentiment counts are included.

• Dataset: Tweets labeled as positive/negative
• Outputs: Sentiment predictions, demo bar plot, question answering
• Validation accuracy: 0.81
• Learned about BERT fine-tuning, evaluation, and inference on new examples

Summary:
Built a deep convolutional GAN from scratch in TensorFlow to generate images of clothing items. The project features a custom training loop, TPU acceleration for fast training, and a streamlined architecture for efficiency. Final results include high-quality generated images and a clear adversarial training dynamic.

• Dataset: Fashion MNIST (28x28 grayscale images)
• Techniques: Custom GAN, mixed precision, TPU acceleration
• Key learning: Adversarial training and performance optimization

Summary:
Used a Random Forest classifier to predict car acceptability from categorical features in the UCI Car Evaluation dataset. The project includes ordinal encoding, stratified splitting, and detailed hyperparameter tuning with validation curves. Visualizations highlight class distribution, validation curves, confusion matrix, and feature importance.

• Dataset: UCI Car Evaluation (categorical features)
• Techniques: Ordinal encoding, Random Forest, validation curves, feature importance
• Key learning: Hyperparameter tuning and feature analysis are critical for model performance

Summary:
Built a robust spam detection pipeline with text preprocessing, feature engineering, and multiple vectorization techniques. Compared Naive Bayes, Logistic Regression, and Random Forest classifiers. Naive Bayes with CountVectorizer achieved 98.7% accuracy on the test set. Visualizations include model performance, confusion matrix, and word frequency analysis.

• Dataset: spam.csv (SMS spam/ham messages)
• Techniques: Feature engineering, CountVectorizer, TfidfVectorizer, model comparison
• Key learning: Simpler models can outperform complex ones for text classification

Summary:
Built and visualized a decision tree classifier to predict the best drug for a patient based on health metrics. The model achieved 97.5% accuracy and provides clear, rule-based predictions. Visualizations include feature importance, confusion matrix, and the full decision tree.

• Dataset: drug200.csv (patient health metrics)
• Techniques: Categorical encoding, decision tree, visualization
• Key learning: Decision trees offer interpretability and strong performance

Summary:
Built and trained a convolutional neural network to classify images of roses, daisies, dandelions, sunflowers, and tulips. The pipeline includes automated data loading, preprocessing, and extensive data augmentation to improve generalization. The model achieved 91.2% accuracy on the training set and 73.6% validation accuracy.

• Dataset: 3,670 images, 5 flower species
• Techniques: Data augmentation, dropout, normalization
• Key learning: End-to-end image classification pipeline and the impact of augmentation

Summary:
Applied RFM (Recency, Frequency, Monetary) analysis and K-Means clustering to segment retail customers into four distinct groups. The project features robust preprocessing, outlier removal, and comprehensive visualizations. Key business insights include identifying high-value, loyal, at-risk, and occasional customers.

• Dataset: Online Retail (4,293 customers, 3 RFM features)
• Techniques: RFM analysis, outlier removal, K-Means, visualization
• Key learning: Translating clustering results into actionable business insights