Experience

• 

  Bjit group

  Aug 2015 - Jun 2018

  [1] eBook Reader: Developed a reader application for smart devices (Andriod & iOS) designed to read and manage e-book contents. Contributed to feature implementation, bug fixing, prototype design, plugin modules, and test case generations.[2] Object Detection: Developed an IoT smart application designed to detect objects and display relevant warnings or advertisements. Utilized Latent SVM OpenCV APIs for image processing, and JobScheduler to synchronize and display contents from servers. Fixed non-compliance issues across different smart TVs, OS versions, and layouts of devices. Show less

  • Software Engineer

    Aug 2015 - Jun 2018

  • ICT Trainer (B-JET)

    Nov 2017 - May 2018

  • Programmer

    Aug 2015 - Feb 2016
• 

  University of houston

  Aug 2018 - May 2024

  [1] LLMs for Code: (a) Trained BERT, GPT, and T5-based pre-trained models using PyTorch and HuggingFace pipelines for various software engineering tasks. (b) Inspected recent code LLMs, such as CodeBERT, CodeGPT, CodeT5, CodeGen, and InCoder, for code generation and defect detection. (c) Probing LLMs and evaluating their robustness, generalizability, memorization, and vulnerabilities for safe coding.[2] Trojans for Code: (a) Designed various backdoor attacks (e.g., variable renaming or dead code insertion) for code models across different tasks, datasets, and programming languages. (b) Generated baseline Trojan models by fine-tuning pre-trained code models with clean and poisoned data. (c) Analyzing model parameters, embeddings, attentions, activations, probabilities, and losses for identifying and unlearning Trojan models and input triggers. Show less [1] Source Code Embeddings: Trained PyTorch and TensorFlow-based deep neural models, such as LSTM, Transformer, Code2Vec, and GGNN, on programming datasets and tasks. Represented programs as sequences of tokens, features, paths, trees, graphs, and images for learning effective vector embeddings.[2] Adversarial Robustness: Introduced a program augmentation technique for evaluating the generalizability and robustness of neural code models. Applied program refactoring, e.g., identifier renaming or reordering, to generate new adversarial programs by inducing semantic-preserving changes to original input programs.[3] Interpretable Prediction: Implemented a program reduction approach to identify important input features that models learn from training datasets for making predictions. Applied program simplification techniques to remove irrelevant parts from input programs while preserving the original predictions of the neural program model.[4] Memorization and Generalization: Evaluated the impact of clean and noisy data on training neural models for various programming tasks. Induced synthetic noise to the training data and studied memorization-generalization trends on the real-world test data, including Transformer-based language models and generative tasks.[5] Analyze Bug Reports: Proposed a clustering-based ML technique to configure test generators for generating effective failure-inducing test programs using bug reports. Evaluated the effectiveness of test configurations for Csmith in finding bugs and assessing coverage in GCC compilers at different optimization levels. Show less

  • Research Consultant

    Jan 2024 - May 2024

  • Postdoctoral Fellow

    May 2023 - Dec 2023

  • Graduate Research/Instructional Assistant

    Aug 2018 - May 2023
• 

  Ul research institutes

  Dec 2023 - now

  Working for a safer digital world! The Digital Safety Research Institute (DSRI) at UL Research Institutes (ULRI) aims to partner with other safety-minded organizations to help protect individuals from rapidly evolving digital threats. More: https://ul.org/research/digital-safety My primary focus centers on the digital safety evaluation of cutting-edge AI tools that assist in programming and software engineering tasks.Website: https://ul.org/research/digital-safety

  • AI Research Engineer

    May 2024 - now

  • Research Engineer

    Dec 2023 - May 2024