• Designed and developed a module to extract bugs from Github repositories and using them as reference to generate defected front-end code with Gemini API to create FCA dataset (Front-end and Code Anomalies dataset 1M+ structured data rows).• Tested various pre-trained model including Llama2 and Mistral v2 with the FCA dataset (Maximum accuracy: 49%).• Fine-tuned Llama2, applying parameter-efficient fine-tuning (PEFT) for optimal resource utilization. Within the PEFT framework, we… Show more • Designed and developed a module to extract bugs from Github repositories and using them as reference to generate defected front-end code with Gemini API to create FCA dataset (Front-end and Code Anomalies dataset 1M+ structured data rows).• Tested various pre-trained model including Llama2 and Mistral v2 with the FCA dataset (Maximum accuracy: 49%).• Fine-tuned Llama2, applying parameter-efficient fine-tuning (PEFT) for optimal resource utilization. Within the PEFT framework, we implemented both Low-Rank Adapter (LoRA) and Quantized LoRA (QLoRA) techniques.• Incorporated Retrieval-Augmented Generation (RAG) to comprehensively capture project functionalities from the codebase, enriching the LLM's understanding.• Implemented Reinforcement Learning from Human Feedback (RLHF) using the HuggingFace trl library to achieve a significant precision improvement. Show less

• Contributed to a 30% improvement in portfolio generation model performance by testing and refining algorithms for portfolio generation using python.• Designed a PyTorch Gym environment for simulating stock trading and portfolio management, utilizing historical stock data like S&P500 to ensure authenticity.• Employing OpenAI’s Proximal Policy Optimization (PPO) algorithm, resulting in a finely tuned trading model.• Calculated and optimized various performance metrics, including… Show more • Contributed to a 30% improvement in portfolio generation model performance by testing and refining algorithms for portfolio generation using python.• Designed a PyTorch Gym environment for simulating stock trading and portfolio management, utilizing historical stock data like S&P500 to ensure authenticity.• Employing OpenAI’s Proximal Policy Optimization (PPO) algorithm, resulting in a finely tuned trading model.• Calculated and optimized various performance metrics, including total return, volatility, max drawdown, information ratio (IR), hit ratio, and win-loss ratio. Show less