CV

BSc. in Data Science and Artificial Intelligence

Hanoi University of Science and Technology, Hanoi, Vietnam

GPU Engineer

Moreh Inc, Seoul, South Korea

• Implement a pure HIP C++ version of OpenAI’s MoE GPT-OSS from scratch (without rocBLAS/hipBLAS); optimize model loading, continuous batching, multi-streaming, multi-GPU communication, CPU–GPU–SRAM memory access, FlashAttention, MFMA GEMM, MoE load balancing; achieved 30k TPS (20B) and 10k TPS (120B) on a single node with 8× AMD MI250x GPUs.

LLM Researcher

Menlo Research, Singapore, Singapore

• Developed a lightweight Speech Tokenizer (22M) using Residual Vector Quantization, achieving SOTA results on viVoice and LibriSpeech by extending the codebook size to 2048, training a two-phase training process (KL+CE and CE loss), contributing to the open-source WhisperSpeech codebase; outperformed original Whisper and PhoWhisper
• Researched the Speechless model, a modified Llama 3.2 1B architecture, to generate synthetic semantic audio representations from multimodal inputs using ASR datasets; paper accepted at Interspeech 2025 (Rank A)
• Modified the Llama tokenizer and performed continual pre-training on semantic tokens from ASR datasets, followed by post-training with mixed raw text, sound-text, and noise sound datasets (filtered by language identification, deduplication, length, and quality) to align with user preferences
• Published the package-modularized Ichigo on PyPI, supporting an asynchronous API for platform developers and implementing audio chunking with overlapping to support long audio input for ASR

Data Scientist

Viettel Group, Hanoi, Vietnam

• Developed a multi-agent Conversational Recommendation System with multimodal capabilities, supporting vision input and speech-to-speech interaction with end-users
• Implemented AdaptiveICL to align with pre-defined expertise plans and designed a synthetic data pipeline for fine-tuning reasoning, SQL query generation, and function calling
• Built Retrieval, Ranking, and Query tools for database interaction; implemented a Candidate Bus for item candidate storage and Web Search for external resource integration
• Engineered an end-to-end system, including a Docker-wrapped API to bridge Application and Infrastructure layers

Organizer

Google Developer Groups, Hanoi, Vietnam

• Planned and created project blueprints, delegated tasks to individuals, connected teams, and provided future operational guidance, certified by Google's Global Headquarters

Applied Scientist

VinBrain (NVIDIA), Hanoi, Vietnam

• Developed MedNeXt for rectal cancer diagnosis and treatment support

Research Assistant

Data Science Laboratory, HUST, Vietnam

• Researched class-specific prompt rehearsal-free in continual object detection

Computer Vision Researcher

Viettel High Tech, Hanoi, Vietnam

• Designed the 3DNeRV architecture utilizing efficient cube-wise embedding for hybrid video representation, enabling parallel frame processing, simultaneous multi-frame reconstruction, and resolution flexibility

Head of Event & PR Team

Google Developer Students Club, HUST, Vietnam

Machine Learning Researcher

LEAN Platform, Southeast Asia

• Developed real-time Conv3D model for concentration recognition in online self-learning platform with 10000 users

Leo

• Architected an LLMOps system for a personal AI assistant; encompassing Data, Feature, Training, Inference, and Observation components, following clean architecture principles
• Implemented an offline pipeline that retrieves data from data services and stores on S3; designed an ETL pipeline to crawl links and perform quality filtering; set up a feature generation pipeline for fine-tuning datasets and creating vector embeddings indexed in MongoDB for Hybrid Contextual Retrieval; and established a training pipeline with evaluation and serving model on HF/AWS endpoints, all orchestrated by ZenML
• Designed an online pipeline featuring an agentic RAG system, served via API using LiteLLM; utilizes summarization and retrieval tools (powered by fine-tuned LLM endpoints and a vector index database), supports Search MCP server, and incorporates observability components through prompt monitoring and RAG evaluation

Gemini Omni

• Developed a real-time web application showcased at Google I/O Extended Hanoi, featuring speech-to-speech functionality, multimodal integration, and RAG for event updates

Google I/O Extended Hanoi Website

• Designed front-end client and admin back-end websites, deployed hosting, and tracked event emails

Detect Cheating in Examination

• Researched and deployed real-time cheating detection solution using Pose3D and VideoMAE for 50-person exam rooms; featured on VTV24, DanTri, HUST, etc