As a visiting scholar RA, I am working with Prof.Federica Sarro and Prof.Mark Harman for the development of self-evolution LLMs on Software Engineering. What we are aiming for is to have a version of self-CI/CD language model that can adapt library updates when doing version changes.
I am currently working with Collins Aerospace for integrating AI into the design workflow, especially considering the process of digitalization. During the process, I developed an AI assistant to help Collins Aerospace to integrate their design results to a RAG database and can be used to inspire their design process.
For the course Deep Learning in TUDelft, I designed the three different assignments for entry-level to master-level deep learning questions. With my expertise in CNN and ML algorithms, I helped the lecturer to prepare the course contents and exam questions.
- Great Bachelor Research Project (Thesis), with grade 8.5
- Java full-stack development, Python Development, Algorithm Design, NP-Hard Algorithm
- Java, Python, Javascript, Scala, C++, SQL, Prolog
- Multimedia Analysis (Image Processing, Signal Processing, Multimedia Processing)
- Machine Learning, Deep Learning (TA experience of course Deep Learning)
- Computer Vision, Large Language Model
- DTU Space for Statellite System and Communication Design
- DTU Computer Science for Medical Image Analysis
- DTU Computer Science for Fintech Application
- DTU Management for innovation with AI
Supervisors: Dr.Maliheh Izadi, Prof.Arie van Deursen, Jonathan Katzy
Course results so far on track for Cum Laude.
| Listening | Reading | Spoken interaction | Spoken production | Writing | ||
|---|---|---|---|---|---|---|
English |
C1:
Proficient user
|
C1:
Proficient user
|
C1:
Proficient user
|
C1:
Proficient user
|
C1:
Proficient user
|
We are trying to develop an AI assistant that can be used to help improve the longtime-existing design procedure inside the company, while maintaining the creativity of all designers.
Showing how the performance of LLMs varies in different programming related tasks when changing the natural language that is used in the code.
Find energy-efficient computer vision solutions in real-world disaster scenarios: utilise the Nvidia Jetson Nano 2GB platform to enhance disaster scene understanding using UAV-based edge devices; strategically optimise the FANet Model through TensorRT, structured pruning, and knowledge distillation, enhancing processing speed and model performance.
Explored facial expression recognition: integrated a Feature Pyramid Network (FPN) and a Residual Network (ResNet) to refine the multi-level feature representation of facial expressions; developed an innovative approach by integrating FPN and ResNet50, which substantially enhanced early-stage training efficiency, accuracy, and adaptability, illustrating the substantial potential of combined convolutional neural network architectures in complex pattern recognition tasks.
Large Language Models are essential coding assistants, yet their training is predominantly English-centric. In this study, we evaluate the performance of code language models in non-English contexts, identifying challenges in their adoption and integration into multilingual workflows.
The primary objective of this challenge is to establish a new and powerful benchmark for the task of removing raindrops under varying lighting and focus conditions. There are a total of 361 participants in the competition, and we ranked 13 in the final testing phase.