I'm an AI engineer based in Hanoi, currently working at Vingroup on AI agents for an in-car virtual assistant. Most of my work sits between LLMs, backend systems, and real-world products: building agents that understand requests, use the right tools, and respond reliably with low latency.
- In-car virtual assistants: building AI agents that understand driver requests and connect language models with vehicle features and services.
- AI agents: multi-agent workflows, intent routing, tool use, memory, structured outputs, and human-in-the-loop flows.
- LLM applications: RAG, MCP integrations, Text-to-SQL, vector search, evaluation, and observability.
- Backend: asynchronous FastAPI services, SSE streaming, conversation state, Docker, and Kubernetes.
- GPU and model engineering: CUDA/C++ kernel optimization, inference profiling, vLLM/ONNX deployment, and work on both NVIDIA and AMD GPUs.
- Robotics: Vision-Language-Action models, humanoid teleoperation, real-world data collection, and inference on the Unitree G1.
| Area | What I did |
|---|---|
| In-car AI assistant | Currently building AI agent capabilities for Vingroup's in-car virtual assistant, with a focus on tool use, reliable responses, and low-latency interaction. |
| Conversational robotics | Built a Vietnamese/English robot receptionist with LangGraph, specialized agents, RAG, and real-time mission generation. |
| Low-latency interaction | Streamed speech and action commands over SSE, then parsed actions as they arrived so the robot could move before the full response was finished. |
| LLM platform | Put Cerebras, Azure OpenAI, and a self-hosted vLLM server behind one interface, deployed it with Docker/Kubernetes, and added Langfuse tracing. |
| Enterprise AI agents | Worked on multi-agent and MCP systems, including Text-to-SQL and backend services for conversation history and user feedback. |
| Large-scale ML | Processed 100M customer records for deep tagging and product matching, making the pipeline 30× faster. |
| Computer vision | Developed an MS lesion segmentation method that reached 91% accuracy while making the base model 35% smaller. |
Most of my day-to-day work lives in private company repositories. These public projects are a sample of what I experiment with outside that work.
My notes and working examples for LangGraph, PydanticAI, web-search agents, and custom tools.
Small MCP server and client examples that show how to expose tools and data to AI applications.
Experiments with training and deploying NVIDIA Isaac GR00T for humanoid robots.
Vision-Language-Action training and inference, LeRobot dataset workflows, and tests on Unitree robots.
Computer vision and graph-based methods for recognizing medicine packaging and matching pill images to prescription information.
Agents & LLMs
LangGraph · LangChain · PydanticAI · MCP · RAG · Function Calling · Text-to-SQL · Qdrant · Langfuse · LiveKit
Backend & production
Python · FastAPI · Pydantic · AsyncIO · REST · SSE · PostgreSQL · Docker · Kubernetes · GitLab CI/CD · AWS ECR
ML, inference & GPU
PyTorch · TensorFlow · scikit-learn · ONNX · vLLM · CUDA · C++ · CuML · CuDF · Nsight Systems
Robotics & embodied AI
Vision-Language-Action Models · NVIDIA Isaac GR00T · LeRobot · Unitree G1 · Teleoperation · Real-world Robot Inference
- I currently work at Vingroup, building AI agents for an in-car virtual assistant.
- I have also built agent systems for service robotics and financial services.
- At Moreh, I optimized model kernels and benchmarked workloads across NVIDIA and AMD GPUs.
- At VinBigData, I worked on medical imaging and driving-behavior models.
- B.Sc. in Computer Science from Hanoi University of Science and Technology — GPA 3.51/4.00.
- Hackathon awards: Champion, HUST Vietnam Logistics Young Talent · Second Runner-Up, ART Hackathon · Third Place, IBM Hackathon.
I'm open to AI Engineer, AI Agent Engineer, LLM/Backend Engineer, and Applied AI roles. I'm especially interested in work where latency and reliability are real engineering constraints, not just dashboard metrics.
📫 xuananhbka@gmail.com · 📍 Hanoi, Vietnam
I like building AI that can use tools and do useful work beyond the chat box.