AI Engineer, LLM Systems Builder, Backend Developer, Full-Stack Engineer
I build production AI systems end to end — from RAG pipelines and agent tool-calling to deployment, telemetry, and cost control. Most AI demos stop at the prompt; I care about what happens after: evals, latency, cost per user, and people who actually book.
About
IT graduate from Thang Long University. For my graduation project I shipped TripNest AI — an AI travel agent running in production with real VietQR payments — and instrumented every turn of it: token cost, tool success rate, conversion funnel.
I like the unglamorous parts of AI engineering: context management, anti-hallucination guardrails, atomic booking logic, and knowing exactly what each user costs per month.
Most recently I interned at VinSmart Future (AI Models & Platform Division), working across two internal AI platforms on agent architecture, plugin and tool safety, deployment governance and LLM unit economics — and reporting the failure modes I found: an evaluation rubric that had saturated, a memory write path that promoted unverified claims, a plugin that could approve itself, and a lost update hiding behind an HTTP 200.
Agents, RAG, evals, guardrails — shipped to production, not just notebooks.
FastAPI services with SSE streaming, atomic booking logic, and telemetry.
Next.js + TypeScript UIs: chat widgets, dashboards, booking flows.
Multi-provider LLM layers, canonical data formats, cache-friendly architecture.
Docker on Fly.io, Vercel, CI with GitHub Actions — deploy and operate for real.
Red-team findings, race conditions, cost blowups — find the root cause, fix it, add a test.
Tech Stack
The stack behind TripNest AI and my day-to-day work — AI-first, but full-stack end to end.
Experience
06/2026 — 08/2026
06/2025 — 05/2026
05/2024 — 12/2024
Featured Projects
Selected work — the flagship is in production with real users and real payments.
05/2026 — 07/2026 · production
Solo developer — design, build, deploy, operate
An AI travel agent for the Vietnamese market, running in production: LLM tool calling across 11 agent tools, RAG over a 96-room knowledge base, and a full booking flow with VietQR payments — instrumented down to cost per user per month.
AI hướng dẫn & kiểm tra hồ sơ thủ tục hành chính
Trợ lý AI cho thủ tục hành chính công: người dân mô tả nhu cầu bằng ngôn ngữ tự nhiên, nhận checklist cá nhân hoá, chụp ảnh giấy tờ để OCR tự điền biểu mẫu, và được kiểm tra hồ sơ trước khi nộp — mọi hướng dẫn đều truy vết được về căn cứ pháp lý.
A grader that never trusts the score the model reports
A meta-agent that grades another agent's output against a 5-criterion rubric. Its central design constraint comes from a measurement: the same LLM, grading the identical criteria vector three times, wrote three different totals — and once omitted a score entirely. So the implementation recomputes every number itself and fails closed.
Sáu bài về đưa agent vào production, mỗi bài kèm mục tự chỉ ra điều có thể sai
Tuyển tập sáu bài phân tích và thiết kế về việc đưa AI agent vào production, mỗi bài bắt đầu từ một sản phẩm có thật và kết thúc bằng một mục ghi rõ điều gì có thể chứng minh tôi sai. Site có agent chạy hàng tuần đi tìm chỗ bài viết đã lỗi thời.
Fine-tuning Qwen2.5-3B with preference optimization
Hands-on alignment training beyond API-level work: fine-tuned Qwen2.5-3B with Direct Preference Optimization and ORPO, comparing preference-optimization strategies on a small open model.
Routing and shared state across specialized agents
Coordinated multiple LLM agents with routing and shared state — patterns for splitting work across specialized agents instead of one monolithic prompt.
Data quality checks and monitoring for ML-ready data
Built a data pipeline with observability baked in — data quality checks and monitoring so bad data is caught before it reaches a model.
E-commerce storefront for PC parts
E-commerce storefront for computer components — built the UI and optimized the product-display flow. Live on Vercel.
GitHub
Live stats pulled from my GitHub profile.
Skills
Honest levels — no fake percentage bars. Proficient means I've shipped it to production.
Education & Leadership
2021 – 2025
B.Sc. in Information Technology
Focus: Information Systems, Data Analysis, Web Applications. Graduation project: TripNest AI — AI Travel Agent (deployed to production).
2024 – Present
Thang Long University
Lead club operations and organize university-wide events.
Contact
Looking for an AI Engineer intern who ships past the demo? I'm happy to walk through the TripNest architecture and telemetry live.