AI R&D engineer at Enclave (Da Nang) and MSc candidate at Da Nang University of Science & Technology. I take LLM features from prototype to production — retrieval that stays grounded, agents that call tools reliably, and the AWS/GCP infrastructure holding it up.
Currently
- Multi-agent orchestration over the A2A protocol — typed tool contracts,
plan → act → observe → iterate - Cutting LLM token cost with response caching, prompt rewrites and per-task model routing
- Backend scalability across enterprise products in NestJS and FastAPI
| Project | What it does | Stack |
|---|---|---|
| Real-estate AI assistant | One orchestration agent routing to focused sub-agents instead of a single overloaded prompt — lower token spend, in-domain tuned model. | ADK-python Gemini Vertex RAG Next.js |
| Agentic RAG job platform | Every reply grounded in retrieved evidence, not free generation. Vector search behind a NestJS gateway. | NestJS FastAPI LangChain Redis Docker |
| Agent-based project planner | Turns raw requirements into structured project plans via coordinated sub-agents. | FastAPI A2A Structured outputs |
| High-availability AWS deploy | Zero-downtime rolling deploys, multi-AZ ECS, least-privilege IAM. | ECS/ECR ALB S3 IAM |
AI OpenAI Gemini Qwen vLLM llama.cpp
Retrieval LangChain Qdrant pgvector Vertex RAG
Backend TypeScript Python NestJS FastAPI Express Redis PostgreSQL
Cloud AWS ECS/ECR GCP Docker GitHub Actions
Have an AI system that needs to survive real traffic? → nhatlinh.dut.3@gmail.com