Machine Learning Research Scientist
Bluelight AI
Total des années d'expérience :0 years, 5 Mois
• Engineering algorithms to automatically detect & repair ML model failures, achieving a 50% performance improvement for specific failure groups & saving the data science team over 20% of their time, & presented results to top clients & investors
• Specializing in training Large Language Models (LLMs), Generative AI models, Retrieval-Augmented Generation models, & Computer Vision models using prominent datasets to analyze performance & investigate model failures. Presented findings to both technical & non-technical audiences including investors & clients
• Developing high-quality production code to build reliable software for analyzing ML models, ensuring robust performance
Multi-evidence Natural Language Inference for Clinical Trial Data Stanford’s Machine Learning Lab Published by The 17th International Workshop on Semantic Evaluation (Link) ● Implemented a system for natural language inference in breast cancer clinical trial reports using a fine-tuned Bio+Clinical BERT Multi-task Learning using BERT Stanford University ● Fine-tuned a BERT model to simultaneously perform 3 tasks - paraphrase detection, semantic similarity & sentiment analysis ● Used gradient surgery, regularization methods & hyperparameter tuning to boost performance & efficiency for all tasks Video Action Recognition Stanford University ● Implemented YoloV7, CNN + BiLSTM, HOG + SVM models on the UFC101 dataset to classify videos over 101 action classes Reinforcement Learning for Robotic Piano Playing Stanford University ● Implemented 2 off-policy reinforcement learning methods (Soft-Actor-Critic & Twin Delayed Deep Deterministic Policy Gradients) to train shadow hands to play the piano in a high-dimensional environment, using Mujoco ● Experimented with data augmentation techniques & hyperparameter tuning to improve model performance & efficiency Efficient ML Systems: Quantization & Pruning of nanoGPT Stanford University ● Employed a nanoGPT for training & evaluation on the Wikitext103 causal language modeling task ● Implemented 8-bit quantized inference & post-training quantization of model weights for a GPT-2 pretrained model ● Executed iterative magnitude based pruning of model weights while continually fine-tuning on the dataset in each iteration Skills: Python (PyTorch, TensorFlow, OpenCV, HuggingFace Transformers, Scikit-learn), C++, Spark, SQL, CUDA, R, Snowflake, Tableau, MongoDB, AWS, GCP, Hadoop, Django, Flask, C, Git ● Extra-Curriculars: Stanford Machine Learning Lab, NUS Statistics Society ● Other Internships: ML Engineer with Dr. Emma Brunskill (Stanford’s AI for Education Lab), Teaching Assistant at Stanford’