Le Duc Anh Tuan

Hi, I’m Charles! Interested in LLM Systems, with experience spanning AI Research (ML, CV, Speech) to Applications (LLMs), and across the stack - from low-level GPU kernel optimization (CUDA, HIP) to high-level LLM system design, training to serving.

Industry:

• Implemented a pure HIP C++ of OpenAI’s MoE GPT-OSS from scratch on AMD GPUs, optimized model loading, continuous batching, multi-streaming, multi-GPU communication, MoE scheduling, CPU-GPU–SRAM memory access, FlashAttention, and MFMA GEMM kernels; achieved 30K TPS (20B) and 10K TPS (120B) on a single node with 8× AMD MI250 GPUs and featured on r/LocalLLaMA.

• Architected Leo - an end-to-end LLMOps system integrating offline and online pipelines for data, training, and inference; built offline pipelines for GET data, ETL, SFT, feature generation, and hybrid RAG ingestion enhanced by a contextual agent, all steps managed by an orchestrator; implemented an agentic RAG-based online system served via API with a React FE and proxy-connected BE, supporting scalable BE services, replica database nodes, vLLM inference engine; incorporated Redis prompt caching, short- and long-term memory modules, CI/CD RAG evaluation triggers (retrieval & generation), and real-time token streaming via SSE.

• Developed a multimodal, multi-agent conversational recommendation system with vision and speech-to-speech interaction; integrated AdaptiveICL, synthetic data generation, retrieval-ranking pipelines and achieved Top 1 in track DSAI at Viettel Digital Talent 2024.

Research:

• Researched Speechless (Ichigo family LLM model), aiming to generate synthetic semantic audio representations from multimodal inputs, trained on semantic tokens generated by Ichigo Whisper; published a paper accepted at Interspeech 2025 (CORE Rank A conference in Speech Processing).

• Proposed “Efficient Continual Detection Transformer”, leveraging pseudo-labeling, knowledge distillation, and LoRA; achieved high performance with only 3% trainable parameters compared to RT-DETR on COCO dataset.

Community:

• Released and maintained open-source projects for developers, including gpt-oss-amd (150 ⭐️), nvims (100 ⭐️), leo, gemino; contributed to leading projects like Ichigo (2.4k ⭐️) and WhisperSpeech (4.5k ⭐️).

• Organized multiple technical events for the developer community, such as VinAI Day, Google I/O Extended, Google DevFest, International Women’s Day x Flutter Forward Extended, and Google Build with AI; spoke at Google DevFest 2022 on “Detecting Cheating in Examinations” (highlighted on VTV24); certified by Google’s Global Headquarters.