Trung Thanh Pham

I am a PhD student in Artificial Intelligence at Kyunghee University Global Campus (Feb 2025 - Present), specializing in quantum computing and Multi-Modal Large Language Models (MLLMs). A product-oriented engineer with robust model training skills and a visionary architecture design mindset. I excel at designing, orchestrating, and implementing best-fit AI approaches to deliver production-ready solutions. I hold a Bachelor’s degree in Artificial Intelligence from FPT University (GPA: 8.87/10, thesis graded 9.2/10).

Research Interests

My research spans Quantum Machine Learning, Large Language Models (LLMs), Computer Vision, and Agentic AI systems. Current and past focus areas include:

• Quantum-classical hybrid architectures — reducing VGG19’s 3M parameters to 150 quantum training parameters with improved accuracy
• Multi-modal LLMs for document understanding, key information extraction, and structured data conversion
• Exemplar-based video colorization using Swin Transformers with temporal consistency
• Deep learning-based 3D reconstruction and single-camera motion capture
• Domain-driven neural machine translation with multi-agent pipelines and low-resource language modeling

Technical Expertise

Languages: Python, C/C++, Java, SQL, JavaScript, ReactJS, HTML/CSS
ML/DL Frameworks: PyTorch, TensorFlow/Keras, ONNX
MLOps & Agentic AI: LangChain, FastAPI, vLLM, Hugging Face TGI, Docker, Azure Pipelines, Weights & Biases
LLM/VLM Fine-tuning & Serving: LLaMA, QwenVL, InternVL, NLLB; closed-source APIs (OpenAI, Gemini)
Core Architectures: Transformers (BERT, GPT, T5, ViT, Swin), CNNs (ResNet, UNet), RNNs (LSTM, GRU), Diffusion Models, GANs, VAEs, Reinforcement Learning, Evolutionary Algorithms (NEAT, GA)
Quantum Computing: PennyLane, TorchQuantum
Cloud & HPC: NVIDIA H200/H100/A100/A6000 GPU server management, Azure AI services, resource allocation for scalable team workloads

Experience

I have worked as an AI Researcher at Kyunghee University and as a Mid-level AI Engineer at IMT Solutions, delivering production agentic systems with 94% peak accuracy and cutting inference costs by over 40%. Previously, I interned at GMO-Z.com RUNSYSTEM, improving deepfake and anti-spoofing F1 scores from 0.86 to 0.91 using Fourier transform approaches.