Mô Tả Công Việc VINFAST is a pioneering electric vehicle (EV) company committed to revolutionizing the automotive industry with sustainable and innovative mobility solutions. As a leading player in the EV market, VinFast is dedicated to delivering high-quality, cutting-edge electric vehicles that redefine the driving experience.
Our team consists of passionate professionals driven by a shared vision of creating a greener and more sustainable future through innovation, technology, and excellence. In this role, you will be instrumental in SLP Center, using your skills to process and analyze large volumes of data. You will collaborate with diverse teams, including Engineering, Quality Control, VINFAST's suppliers, to create cutting-edge solutions that will drive the future of transportation.
Research, design, and train foundation language models (LLM/SLM/Omni) powering robot and virtual assistant products, spanning the full lifecycle from pretraining through post-training.
Build and optimize production-grade LLM agents, voice agents (accuracy, latency, scalability, reliability) on top of state-of-the-art foundation models (OpenAI, Gemini, Claude, open-source LLMs, Finetuned In-house), including model evaluation and selection, multi-agent orchestration, planning, memory, tool calling, workflow execution, voice interaction handling and RAG pipelines (retrieval, reranking, indexing, embedding optimization, knowledge integration).
Research and develop omni models, full-duplex speech-to-speech models, and turn-taking management models for full-duplex voice agent experiences.
Build data pipelines for preference/instruction/agentic data collection and processing (data curation, synthetic data generation, data mixing, deduplication, quality filtering).
Own the end-to-end LLM training pipeline: pretraining (data curation, tokenization, dense/MoE architectures, scaling laws, distributed GPU training), domain-adaptive pretraining for Vietnamese, dialogue and robot/virtual assistant use cases, and post-training/alignment (SFT, RLHF/RLAIF, DPO/PPO/GRPO, reward modeling, constitutional AI, safety tuning, hallucination reduction).
Design and implement model compression, quantization and distillation methods to produce efficient SLMs for on-device/edge deployment (robots, virtual assistants).
Optimize training performance: distributed training (data/tensor/pipeline/expert parallelism), mixed precision, gradient checkpointing, and other memory/compute-saving techniques.
Build evaluation pipeline across training / inference stages: internal benchmarks, automatic evaluation, human evaluation, A/B testing and safety/alignment test suites.
Track the latest research in pretraining/post-training (papers, open-weight models, training techniques) and rapidly prototype and integrate promising methods into production pipelines.
Collaborate with QC, BA, and Infra teams to ensure foundation models meet production deployment requirements (quality, latency, cost, scalability).
Mentor junior engineers, conduct technical reviews, and contribute to engineering standards for the team
Yêu Cầu
Bachelor's or master's degree in computer science, Artificial Intelligence, Machine Learning, Data Science,
Software Engineering, or related field. - 3+ years of experience training/fine-tuning large language models, or deep research experience in deep learning/NLP/LLM. (Adjust based on seniority.) - Strong understanding of Transformer architecture, attention mechanisms, positional encoding, and modern architectural variants (MoE, SSM, hybrid architectures). - Hands-on experience with LLM pretraining from scratch, continued pretraining, or finetuning including data pipeline design, tokenizer training, and scaling. - Hands-on experience building AI agents using frameworks such as LangGraph, LangChain, LlamaIndex, Semantic Kernel, AutoGen, CrewAI, or equivalent. - Proficiency in post-training techniques: SFT, RLHF/RLAIF, DPO, PPO, with clear understanding of the tradeoffs between methods (offline vs. online RL, reward-based vs. reference-free). - Experience building reward models and preference data collection pipelines (human or AI feedback). - Strong Python skills and experience with distributed training frameworks: PyTorch, DeepSpeed, FSDP, Megatron-LM, or equivalent. - Experience with training/fine-tuning libraries: LLamafactory, Hugging Face Transformers, TRL, PEFT, Axolotl, or equivalent. - Understanding and hands-on experience with parameter-efficient training techniques (LoRA, QLoRA) as well as large-scale full fine-tuning. - Experience working with multi-node/multi-GPU infrastructure, optimizing communication overhead and memory usage. - Strong analytical skills, ability to read research papers, and rapidly translate them into experimental pipelines. - Strong communication skills and ability to work cross-functionally with other teams.
Quyền Lợi
Attractive Compensation and Benefits: Competitive income package, including a 13th-month salary, performance bonuses, and other reward programs in accordance with Vingroup's policies. - Healthcare: Access to high-quality health insurance, ensuring safety and comprehensive employee welfare. - Working Environment: A modern, civilized, professional, and safe workplace offering ample opportunities for personal development and career advancement.