Today
I’m an AI Software Engineer and Tech Lead at FPT Software in Hanoi. I build multi-agent LLM platforms — workflow engines, document AI, and the observability and governance that make them dependable in production.
Previously a software engineer at SETA International, working on high-performance React applications. NVIDIA Certified Associate in Generative AI and LLMs.
Projects
Multi-Agent AI Platform / observability & governance Making a production multi-agent system legible — end-to-end tracing, cost accountability, and the governance layer leadership actually uses. Enterprise Document AI Assistant / the ingestion backbone The document-processing spine of a five-microservice AI assistant — SharePoint Q&A, tabular analysis, translation, and image generation via LangGraph agents. LLM Agent Workflow Platform / AgentflowV2 A low-code LLM workflow builder on Azure — and its second-generation graph engine with loops, conditionals, and live execution tracking, shipped to GA.
Open source
kimi-skills · ★ 230 AI agent skills for professional document generation (DOCX, PDF, XLSX) with native charts and full OpenXML control. free-tts · ★ 3 Open-source text-to-speech web app that turns text into natural-sounding speech. eml-extractor · ★ 2 Pure Python library for extracting headers, body, attachments, and inline images from .eml files — no external dependencies. next-marketing-site · ★ 3 Starter template for marketing sites with full SEO support — Next.js, TypeScript, Tailwind CSS.
Also a contributor to Onlook (23.6k ★) — the open-source Cursor for designers — with four merged pull requests.
Writing
Decision records are a product surface The best architecture notes do more than preserve context — they make the next decision easier. When not to add an agent A small rubric for deciding whether a workflow needs autonomy or simply needs a clearer interface. Ship the first vertical slice Why the fastest route through ambiguity is usually a complete, narrow path through the system.