GIORGI
GOGITIDZE

( SOFTWARE ENGINEER )✶BUILDING SCALABLE SYSTEMS

Giorgi Gogitidze
Scroll Down

WHO / WHAT

Passionate about building scalable, beautiful web experiences and researching NLP & reinforcement learning. Full-stack with React, Go, Node.js, Python. Always open to new opportunities and collaborations!

Capabilities
  • Product Design
  • Full-Stack Development
  • Systems Thinking
  • User Research
  • AI / ML Integration

Experience

The Cog AI Lab, Caldwell University
03/23 - 05/25

Machine Learning Research Assistant

Developed and visualized model performance, improved GPT-2 training speed by 20%, reduced response time by 23%, and increased lexical diversity through reinforcement learning integration.

  • Leveraged Python, TensorFlow, Keras, and Matplotlib for model transparency and interpretability.
  • Published research: Using a GPT with Reinforcement Learning for More Personalized or Targeted Text Generation.
PythonTensorFlowKerasReinforcement Learning

Technical Arsenal

JavaScript
TypeScript
React
Node.js
Go
Python
MongoDB
PostgreSQL
Docker
Git

GPT-2 + Reinforcement Learning for Targeted Text

CCSCNE 2024 poster on applying conservative PPO updates to a frozen GPT-2 policy. Manual reward shaping penalized repetition, rewarded lexical diversity, and raised diversity from 0.499 to 0.616 while reducing repetition by 23%.

Tech Stack
GPT-2 (124M), Proximal Policy Optimization, Manual reward shaping
Role
Lead Researcher & Presenter
Year
2024

LSTM Modeling for Technical Stock Analysis

Caldwell University Research & Creative Arts Day study comparing an MSFT forecasting LSTM to linear baselines. Sliding-window preprocessing stabilized training and drove MAE down from 3.65 at epoch 100 to 3.09 by epoch 500.

Tech Stack
LSTM, Sliding window preprocessing, MAE / MSE evaluation
Role
Research Co-Author
Year
2023

Non-Euclidean Navigation in Reinforcement Learning

Jun-Aug 2022 paid study on training Delayed Q-learning agents inside a Unity-built hyperbolic environment. Documented partial policy transfer from Euclidean layouts and the predictable degradation as curvature and geometric distortion increase.

Tech Stack
Unity Hyperbolic Environment, Delayed Q-Learning, Geometry-aware reward shaping
Role
Summer Researcher
Year
2022