Bongkeun Shin

• RAG Evaluation Architecture Project: Research to evaluate the performance of RAG. • Critique RAG Project: Designed and developed a new RAG architecture improving the baseline paper, Self-Reflective-RAG that increased by 5~10% in accuracy, citation recall, and by 20% in EM(Exact Match) • Topic Modeling Project: Implemented a BERTopic module to Samsung Electronics internal search engine which compared search keyword and retrieved documents. Finetuned BERTopic and increased by 3% in click-hit ratio.

• Document Classification Project: Implemented a classification model by finetuning Sentence Transformer to boost the score of the documents from the search engine • Developed log collecting program to provide data for training and developed log reindexing Python CLI to reindex Elasticsearch indices • Researched 3D reconstruction models and implemented baseline models of NeRF, Nerfies and NSVF for future work. Reproduced the figures such as PSNR, SSIM, LPIPS equal to or greater than ones in the papers.

• HackingZone Web Platform Initiation Project: Designed and developed the bug bounty platform. Implemented virtual desktop accessible in web browsers for white hat hackers to conduct diverse attacks on demo products. Focused on the security to withstand cyber attacks • Conducting sales to customers. Gathered requirements and feedback from white hackers, corporate security managers and C-level executives. • Our team successfully launched a startup company, FindTheGAP, with funding from Samsung SDS

• Developed ERP system with ABAP language for Samsung C&T, Samsung Heavy Industries • Implemented and configured new business process on Sales & Distribution, Logistics Execution for Samsung C&T.

Final Project: Comparing both cost-free and paid models for Chatbot in experiments of finetuning pretrained Llama-2-13b, flan-t5 and GPT-3.5, and conducting Retrieval Augmented Generation on finetuned Llama-2-13b and GPT-4 models. Verified Llama-2 can perform comparable to GPT-4 on RAG experiments