NST

AIENGINEER

Nguyen Son Tung
// grounded answers, not guesses
not vibes
// I'm Tung — an AI engineer building RAG
pipelines & agent harnesses,
shipped to production. Open to projects.
HarnessRagAgenticMCPRedisPyTorchAzureHarnessRagAgenticMCPRedisPyTorchAzure

01 · About

Ho Chi Minh City, VN
AI Engineer · FPT Software
Open to projects

I build AI agents that
do the boring things
so humans don’t have to.

I'm an AI Engineer at FPT Software with 3+ years building production LLM, RAG, and multi-agent systems for enterprise clients. My core specialty is agent harness engineering — designing the tools, sandbox, permissions, and connector layer an agent session runs on — proven on a 20-person production platform where I've been a top contributor. I'm also a daily power user of Claude Code, Codex, and Cursor for agentic development workflows.

02 · Skills

What I build with

Agent & LLM

RAGLangChainLangGraphMCPOpenAI APIAzure OpenAIAnthropic ClaudeHugging FacePrompt Engineering

Agent Harness

OpenSandboxHITL PermissionsAgent MemorySkills Systems

Backend & Data

FastAPISQLAlchemyCeleryPandasPostgreSQLMySQLSQLiteRedisAzure Cognitive SearchBlob Storage

Computer Vision

YOLOPyTorchCNNTransformersCM-GANDeepFillv2ArcFaceTriplet Loss

Cloud & DevOps

AzureDockerKubernetesAzure DevOpsGit/GitHubPytestLangfuse

AI Coding Tools

Claude CodeCodexCursor

03 · Experience

Recent work

  1. ITOCHU GPT

    Now

    AI Engineer

    Multi-Agent AI PlatformClient: ITOCHU CorporationTeam of ~20
    • Own the File Agent subsystem — PPTX/DOCX/XLSX generation inside isolated sandboxes.
    • Delivered video-analytics Q&A and streaming text-to-speech capabilities integrated into the platform.
    • Collaborated across backend, frontend, AI, and infrastructure teams to bring agent capabilities from architecture through production deployment.
    Jul 2025 — Present

  2. Infinity

    Now

    AI Engineer

    Agent Harness & Sandbox PlatformIndependent codebaseTechnical Lead
    • Built the agent harness end-to-end: runtime, tool execution, permissions, and sandbox orchestration.
    • Integrated OpenSandbox to isolate LLM-generated code and file-processing workloads.
    • Designed a human-in-the-loop tool-approval system with pause, persist, and resume.
    May 2026 — Present

  3. AI Buddy

    AI Engineer

    Client: KPMGTeam of 5
    • Built an automated crawling pipeline extracting financial data from structured & unstructured sources.
    • Developed the knowledge base backing real-time RAG querying for the chatbot.
    Feb 2025 — Apr 2025

  4. Ivy Chatbot

    AI Engineer

    FPT Software IvyChat product line
    • Optimized document chunking to improve retrieval accuracy across multiple source types.
    • Implemented LangGraph to structure multi-turn conversation flow and maintain context.
    Dec 2024 — Mar 2025

04 · More Projects · 8 total

Everything else I’ve shipped

2026

Infinity

Agent harness & sandbox platform — runtime, tool execution, permissions, and human-in-the-loop approvals.

Agent HarnessOpenSandboxTool Permissions

2025

AI Buddy

Real-time financial-insights chatbot for KPMG over a continuously crawled knowledge base.

Azure OpenAILangGraphPostgreSQL

2025

Ivy Chatbot

Multi-source conversational assistant — part of FPT's IvyChat product line.

LangChainLangGraphFastAPI

2025

BP Risk Search Agent

LangGraph + GPT + web-crawling agent for automated partner compliance-risk search.

LangGraphFastAPIPostgreSQL

2024

QaiDora Mask Ads Defender

Real-time CV system detecting and blurring prohibited imagery during live broadcasts.

YOLOv8/v9CM-GANOpenCV

2024

Query Vision

LLM system summarizing and answering questions over live AI-camera feeds.

Azure OpenAIPostgreSQL

2023

Vietnamese Sign Language Translator

University capstone — CNN + Transformer model translating sign language to text/speech.

PyTorchMediapipe

—

Helmet Violation Recognition

Detects helmetless motorcyclists using YOLO, PaddleOCR, and ByteTrack.

YOLOPaddleOCRByteTrack

05 · Work

Proof, not
adjectives.

Work that exists outside a CV bullet and can be checked: a filed US patent application.

Years in production LLM/AI
3+
Person platform team
20
Projects shipped
7
US patent filed
1
  1. 01PATENT

    Query Vision

    LLM system answering questions over live AI-camera feeds. Underlying method filed as US patent application US20260189519A1.

06 · Agent stack

How I think about this stack

An LLM is the reasoning engine. An agent adds a loop and tools. A harness gives it the runtime to work safely, statefully, and at scale — the mental model behind everything in Experience and Work above.

01 · Model foundations

Data → pretraining → a base LLM.

Before an agent can act, there is a model: data becomes tokens, a Transformer learns to predict the next one, and pretraining produces a foundation LLM.

  • Data + tokenization
  • Transformer architecture
  • Next-token prediction
  • Base / foundation LLM
Diagram showing raw data flowing through tokenization, Transformer architecture, and pretraining into a base foundation LLM.

// click a layer · arrow keys to step through

07 · Education & Certs

Education

Master of Software Engineering (MSE), Artificial Intelligence

FPT UniversityIn progress

2026 — Present

B.Sc. in Artificial Intelligence

FPT UniversityGPA 8.45/10

2021 — 2025

08 · Contact

Let's build something grounded.

Open to AI Engineer / LLM Application roles. I usually reply within a day.

Available for new work

Best way to reach me

Email

nstung463@gmail.com

Write
Read the résuméOne page, kept current0335 955 790Ho Chi Minh City, VietnamGitHubLinkedInInstagram