Hasan Kemal Abamor - CV

Personal Statement

I am a 23-year-old computer science engineer with a passion for combining practical engineering with innovation. My interests span machine learning, control theory, and explainable AI, and I constantly strive to merge theoretical concepts with real-world applications.

Education

B.Sc. Computer Science and Engineering
Institution: Isik University, Faculty of Engineering and Natural Sciences
Graduation: January 2025

CSE564 Deep Learning Project – Master-Doc Level Course (2025)

• Project Goal: Innovative, first-of-its-kind trial to build a recursive training loop for classifying multiple types of logical fallacies in text, addressing class imbalance and ensuring interpretability.
• Model Architecture:
  • T5 Encoder with MLP Head:
    Backbone: T5 for an encoder-decoder architecture capturing rich semantic representations.
    MLP Head: Multi-Layer Perceptron appended to the encoder output for fallacy classification.
    Layer Freezing: Early encoder layers remain frozen to retain broad linguistic features.
    Input Encoding: Tokenized text mapped into model-friendly vectors.
• Data Augmentation with BLOOM:
  • Synthetic Sample Generation: A pre-trained BLOOM model produces additional training examples for underrepresented classes.
  • Controlled Prompting: Example – “Generate a detailed example of the ‘False Dilemma’ fallacy.”
  • Class Balancing: Newly generated data is integrated into the training set, strengthening performance on minority classes.
• Recursive Training Strategy:
  • Epoch-based Augmentation: Triggered at Epoch 4 once training loss stabilizes.
  • Ongoing Retraining: Model retrained with both original and synthetic data, monitored via training/validation loss.
• Technical Details:
  • T5 Encoder & MLP: Encoder’s last_hidden_state feeds into a two-layer MLP (512-unit hidden layer, ReLU, 30% dropout).
  • Resilient Backpropagation (RPROP) for stable convergence.
  • Explainability: SHAP analysis to identify which tokens/features most influenced classification outcomes.
• Outcome: Demonstrated feasibility of recursively augmented approach to fallacy classification, addressing class imbalance and enhancing interpretability.

Bachelor’s Thesis – Comprehensive Blood Analysis System

Supervised by Tuğba Erkoç

• Developed a non-invasive diagnostic system using Near Infra-Red (NIR) spectroscopy data to predict blood chemical compound levels.
• Implemented an epsilon filtering technique to mitigate overfitting by eliminating redundant data from high-dimensional spectral measurements.
• Tuned epsilon thresholds by monitoring training vs. validation loss differences, ensuring accuracy improvements generalize.
• Final model: A stacking classifier ensemble (Gradient Boosting, MLP, Logistic Regression) reaching ~72% accuracy in a Zindi competition.
• Prepared RAD, ODD, and SDD documents, adhering to MVC principles.
• Deployed system as a scalable web application using Python, Flask, and Digital Ocean infrastructure.

Start-up Co-founder – Health Technology Project

• Organizations: TUBITAK BIGG, Bogazici University (2023)
• Ranked among the top eight teams in Turkey for startup innovation.
• Collaborated with co-founder Cem Dogan (20+ years in electronics) to develop medical tech integrating advanced ML and sensor solutions.
• Designed scalable systems using a modified spectrophotometer (HAMAMATSU) for detecting cardiovascular diseases.

Research Collab – EEG OCD Data Analysis

Institution: Isik University, 2024

• Assisted with ML implementation for EEG data analysis focusing on identifying OCD patient patterns.
• Explored advanced mesh modeling of brain layers, signal processing, and source prediction under Rüştü Murat Demirer.
• Utilized MATLAB and PyCharm for analysis.

Professional Experience

Data Scientist Intern – SHAP Implementation for Big Data (2024)

Company: Datamind (IBM Gold Partner)

• Optimized SHAP Computation: Built a Dual-DB, Multi-Threaded SHAP Calculation System for high-dimensional data, near real-time feature attribution.
• Integrated Redis for ultra-fast lookups, drastically reducing query response times.
• Leveraged SIMD, parallel execution, and batched processing for a 10x speedup over standard SHAP methods.
• Improved stability with optimized garbage collection and memory management.
• Enhanced query performance with PostgreSQL indexing, partitioning, and parallelization.
• Cut runtime from hours to seconds via multi-threading and batch processing in model execution pipelines.
• Built performance monitoring to track CPU, RAM, disk I/O, and network usage, removing bottlenecks.
• Developed scalable API endpoints for interactive SHAP analysis and visualization.

Certifications

• AI and High Performance Computing (HPC): Hewlett Packard Enterprise (Completed)
• Foundations of Cybersecurity: Google (Completed)
• Introduction to Concurrent Programming with GPUs: Johns Hopkins University (In Progress, 50% complete)

Skills

• Programming Languages: Python, MATLAB, SQL, Java, Bash, C
• Machine Learning: Explainable AI (SHAP), Transfer Learning (Ensemble Learning)
• Control Systems: State-Space Modeling, Transfer Functions, System Identification
• Mechanical Design: Temperature Control Systems
• Databases: PostgreSQL, Redis, MongoDB, InfluxDB, MySQL
• Creative Skills: Coffee Brewing, Flavor Profiling, Competitive Brewing Techniques, Coffee Roasting

Google Docs Version