• Integrated Medusa speculative decoding via TensorRT-LLM in Dobby, achieving a 1.8x Throughput improvement for the Vicuna7B model, and successfully benchmarked its performance on Triton Server.• Performed grid search to optimize parameters for the Whisper model, enhancing Word error rate.• Optimized GPU event tracking and usage hours from logs through Python scripts, enhancing data processing efficiency.

• Contributed to multiple guardrails, tests and metrics for the raga llm hub package, for evaluating LLM’sperformance.• Fine-tuned Yolov8 segmentation model, and conducted failure mode analysis to meet customer POC specifications.

• Developing a pipeline to facilitate code-switch (English-Malay) text generation for conversational domain. • Leveraged the Self-Bleu metric for prompt quality assessment and applied the multi-lingual BERTScore toidentify high-quality code-switch candidates.• Research Advisor: Dr. Chng Eng Siong

• Migrated code from SAP UI5 to React framework of OneSAP Portal, an application designed to enable customer success solutions.• Implemented the components required for ticket-raising and live attachments functionality within the application.

Conducted a research under the supervision of Professor, Dr. Puspak Bhattacharya, on developing an automated method for multilingual wordnet linking using sense embeddings.

This project explores the use of image analysis in a two-dimensional spectrum to detect RFI (Radio Frequency Interference) in large amounts of memory-distributed radioastronomical data. As a result, this project aims to supports and develops 1D and 2D memory-distributed convolution within the Heat domain decomposition (or data distribution) scheme.