AI Engineer · Software Engineer · Backend & Applied AI

Building production-ready AI
and reliable software systems.

I design and ship applied AI products, backend services, and developer-focused systems from research prototypes to deployed software. My work spans LLM pipelines, retrieval systems, scalable APIs, and full-stack delivery with a focus on clarity, reliability, and measurable impact.

Based in
Thu Duc, Ho Chi Minh City
Currently
AI Researcher · AI VIET NAM
Available for
AI / Software Engineer roles

A backend engineer with a research habit.

I'm a final-year Software Engineering student at FPT University, majoring in Backend Development. I lead my own AI projects end-to-end — from data curation and model fine-tuning to API delivery, containerization, and evaluation.

I care about systems that are clear, reproducible, and useful in production. Lately I've been working on adaptive reasoning for Vietnamese vision-language tasks, an RAG research platform for Vietnamese retrieval ablations, and small-scale LLM training from scratch. I previously interned at FPT Software building Java/Spring Boot services.

My quiet build companions.

Two small reasons to step away from the keyboard, reset, and come back with a clearer head.

Selected work

Accepted · AICI 2026 Springer edited book

Difficulty-Aware Adaptive Reasoning for Vietnamese VQA with GPT-OSS

Dai-Nhan Tran, Phuc-Thinh Nguyen, Tue-Anh Vu, Bao Do, Hai-Au Trinh, Anh-Khoi Nguyen

We propose an adaptive, difficulty-aware reasoning framework for Vietnamese Visual Question Answering. The system leverages dense captioning from Gemini 2.5 to enrich context for downstream LLMs (GPT-OSS, Qwen3, DeepSeek), with a router that scales inference compute by question difficulty. The framework reaches competitive BLEU@4 / ROUGE-L / METEOR on ViVQA-X while keeping inference budgets low.

  • Venue — AICI 2026, Hanoi (accepted for presentation & Springer chapter)
  • Datasets — ViVQA-X, OpenViVQA
  • Stack — PyTorch · Transformers · vLLM · Unsloth · Gemini 2.5
Accepted · ICISN 2026 Co-author

Curating Multi-Mode CoT for Efficient Math Reasoning with GPT-OSS

Hai-Au Trinh, Tue-Anh Vu, Dai-Nhan Tran, Uyen Khoi-Minh Huynh, Anh-Khoi Nguyen

A two-step chain-of-thought curation pipeline for distilling math reasoning from a GPT-OSS teacher into a smaller Llama 3.2 3B student. We generate multi-mode reasoning traces at low / medium / high inference budgets, then apply answer verification and length-based filters — emphasizing quality over quantity to improve sample efficiency for SFT and GRPO.

  • Venue — ICISN 2026 (accepted)
  • Benchmarks — GSM8K · MATH500 (0-shot & few-shot)
  • Best result — GSM8K 0.8006 · MATH500 0.4760 (Llama 3.2 3B + GRPO)
  • Stack — LLaMA-Factory · MS-SWIFT · GPT-OSS · Llama 3.2

Where I've worked

  1. Oct 2025 — Present Now

    AI Researcher & Teaching Assistant

    AI VIET NAM

    Mentor learners through hands-on labs in NLP and multimodal AI. Contribute to experimentation workflows for Vietnamese VQA and language understanding, and document reproducible research practices for team execution.

  2. Jan 2025 — Apr 2025

    Web Developer Intern

    FPT Software HCM

    Built and scaled Java / Spring Boot backend services backed by PostgreSQL and Redis. Delivered dynamic web features with Thymeleaf, containerized services with Docker, and shipped in an Agile sprint cadence.

  3. Oct 2023 — Oct 2025

    Club President

    FPT Information Technology Club (FCODER)

    Led the student technology community: organized coding events, ran technical workshops, and grew peer mentoring across cohorts.

Things I've built

A mix of research code, full-stack products, and tooling. Pulled from github.com/hugebenevolence.

ViVQA-GPT-OSS-DRA

Research

Reference implementation of our AICI 2026 paper on difficulty-aware adaptive reasoning for Vietnamese VQA. Routes questions across multiple LLM backends based on estimated difficulty.

Python · PyTorch · Transformers · vLLM · Unsloth

View on GitHub →

RAG-Enhancement (RAG Lab)

Research

Full-stack research platform for benchmarking Vietnamese retrieval strategies — BM25, hybrid, RRF, HyDE, MMR, cross-encoder, semantic — with FastAPI workers, Redis queues, and a Next.js comparison UI.

FastAPI · Next.js · Redis · PostgreSQL · Docker

View on GitHub →

vietnamese-gpt2

Research

A Vietnamese language model built on the GPT-2 architecture with a multi-phase training pipeline over crawled web data. Covers tokenization, curation, pretraining, and evaluation for Vietnamese generation quality.

Python · PyTorch · Transformers

View on GitHub →

LLaMA-OSS

Research

Curating multi-mode chain-of-thought data for efficient math reasoning with GPT-OSS. Explores how reasoning style and difficulty interact with sample efficiency at training time.

Python · PyTorch · CoT data curation

View on GitHub →

FLM — Course Management

Internship

Web-based course management system shipped during the FPT Software internship. Modules for to-do tracking, feedback handling, and admin operations; Redis caching and Cloudflare-fronted deployment.

Spring Boot · PostgreSQL · Redis · Cloudflare · Docker

Team Lead

Brainify

Capstone

Team collaboration platform built on a .NET 8 microservices architecture. Designed the service boundaries, API contracts, and inter-service data flow; led the team through planning and delivery.

.NET 8 · ReactJS / Next.js · SQL Server · Redis · Docker

Team Lead & Full-Stack Developer

What I reach for

AI & ML

  • Python
  • PyTorch
  • Transformers
  • vLLM
  • Unsloth
  • LangChain
  • RAG
  • VQA

Backend

  • Java
  • Spring Boot
  • .NET 8
  • FastAPI
  • PostgreSQL
  • Redis
  • REST APIs

Web & Delivery

  • Next.js
  • React
  • TypeScript
  • Docker
  • Cloudflare
  • Vercel
  • Supabase
  • Git

FPT University · Oct 2022 — Jan 2026

B.E. Software Engineering

Major in Backend Development — GPA 3.2.

Capstone in microservices & team collaboration platforms.

Awards & certifications

  • Top 14 — FPT Hackathon 2024 (Donald, AI support for children with autism)
  • President — FCODER, FPT IT Club (2023–2025)
  • Coursera — Software Development Lifecycle
  • MOS — Word 950, Excel 975

Let's build something useful.

Open to AI Engineer and Software Engineer roles, research collaborations, and applied ML problems involving Vietnamese language or vision.