Hire Ran X. - Fine-Tuning Engineer

Full-stack participation in a computer vision (CV) project for law enforcement, including research initiatives, data acquisition, model improvement, auxiliary pipeline development, and product deployment and maintenance at the client site. Supported internal and client reporting with customized tabular and graphic formats at each stage. Led the development and deployment of a video activity detection pipeline, integrating image and video components into the company's media processing product for client-site deployment. Co-led enhancement experiments on an image transformer model. Drove significant improvements and exceeded the desired performance level. Co-led a client-site annotation project, customizing and deploying annotation tools, training government agencies, and ensuring quality assurance. Contributed to research and presentations on alternative approaches to age estimation and activity detection. Assisted in the development of a novel annotation schema for law enforcement use cases, contributing to training material across various formats. Acted as an on-site representative, facilitating communication between government clients and development teams to drive project improvements.

Research assistant in applying Explainable AI (XAI) to Natural Language Processing (NLP) models, and the co-presenter of the final research project for the company's research committee. Design and implement a series of experiments using XAL tools (SHAP and LIME) in sentiment analysis with a pretrained and fine-tuned BERT transformer model. Derived a key phrase extraction pipeline using XAL tools based on sentiment relevancy for explainable model inference interpretation.

* Led the design and launch of an experiment analyzing technical requirements from multi-industrial job post dataset: * Prepared the dataset with UBIAI annotation tool, and experimented with various feature engineering and vectorization methods using Regex, NLTK, SpaCy, and Gensim. * Developed a model that classifies short phrases by training and evaluating the Recurrent Neural Network (RNN) models with Word2Vec embedding, Long-Short-Term-Memory (LSTM), and Conditional Random Field (CRF) layers using TensorFlow and Keras. Also fine-tuned a SpaCy NER Transformer model for performance evaluations. * Experimented with the effect of using Variational Autoencoder (VAE) generated instances in model training by injecting them into the original dataset.