about.md

Nguyen Chi Trung

Software Engineer

I like systems that reveal themselves slowly: data flows, failure modes, and the small decisions that explain the whole.

What I work on

I build backend services, AI-assisted products, and the interfaces that make them usable. My recent work centers on retrieval-augmented answers, role-aware web applications, and the testing and delivery work that keeps them dependable.

• AI and retrieval: grounded answers with source citations, document ingestion, vector search, and model integration.
• Backend systems: API design, authentication and authorization, relational data, background work, and testable architecture.
• Product surfaces: React and Next.js interfaces that stay calm when the system underneath is complicated.
• Delivery: automated tests, CI/CD, code quality checks, and documentation that another engineer can actually use.

How I work

I start by tracing the data end to end. Then I look for the boundary where behavior becomes surprising. I prefer small, verified steps over confident guesses, and I treat tests and documentation as part of the interface rather than cleanup.

At FPT Software, I’m an AI Engineer Intern leading a seven-person team on a syllabus-grounded Q&A agent. The work spans a FastAPI backend, a Next.js frontend, retrieval over course material, and the CI and test coverage around both repos.

The loop

The useful question is usually one layer deeper than the one on the screen.

Off the clock

Outside work, I take systems apart for fun: binaries, protocols, and the small decisions that explain the whole. It rewards patience and careful observation, but it is a personal interest rather than the job I am building toward.

experience.md

Experience

A timeline from school to building AI-assisted products, with the through-line being systems work: understand the whole, then improve the part that matters.

Trajectory

Current role

AI Engineer Intern · FPT Software

06.2026 — present · Team lead / 7

I lead a seven-person team building a syllabus-grounded Q&A agent for course material. The product combines a FastAPI backend, a Next.js frontend, retrieval over documents, and source-cited answers.

My work covers architecture and integration, backend behavior, role-aware access, testing, and delivery. The team ships in short cycles with a Jira board, shared API contracts, and automated checks on both repositories.

Focus map

Timeline

Alongside

Independent Luau scripting

05.2026 — present · Commissions

I build menus, automation, and small tools for Roblox projects. The work is often reverse-engineering adjacent: read an unfamiliar system, find its seams, then write the smallest script that changes behavior cleanly.

Education

X University

09.2023 — 06.2026 · Undergraduate, degree awaiting conferral

Information technology foundations, software engineering practice, and a long run of team projects. The degree is awaiting conferral; the placeholder school name is intentional.

projects.md

Projects

Four systems that show the range: a private AI product, two public .NET applications, and the frontend that makes the agent usable.

Project map

At a glance

Course Q&A

A syllabus-grounded assistant that answers course questions from supplied material and keeps citations beside the answer. The system is split into two repositories: a FastAPI backend and a Next.js frontend. Both are private; access can be requested by email.

• course-qa-agent-be.md — RAG pipeline, Gemini, Qdrant, PostgreSQL, role-aware API.
• course-qa-agent-fe.md — App Router shell, chat surface, citations, shared UI primitives.

devflow-platform

A Jira/Linear-inspired project management platform built as a long-running architecture and DevOps exercise. It covers workspaces, projects, sprints, Kanban boards, dependencies, time tracking, realtime updates, GitHub integration, analytics, and an AI planning agent.

Open the deep dive or visit the live app.

KCPM

A waste-recycling platform built for a software verification course: citizen reports, collector jobs, enterprise operations, and admin review. The project is as much about test discipline and delivery as it is about the product surface.

Open the deep dive or visit kcpm.vercel.app.

Other work

Earlier and smaller builds include net-twin, Smart Nurse, office3d, NUTRI-FIT, and computer-vision experiments. They are not the main story, but they explain the range.

Project index

{
  "primary": [
    { "name": "Course Q&A", "repos": 2, "visibility": "private" },
    { "name": "devflow-platform", "repos": 1, "visibility": "public" },
    { "name": "KCPM", "repos": 1, "visibility": "public" }
  ],
  "live": ["devflow-platform", "KCPM"],
  "focus": ["RAG", "architecture", "testing", "delivery"]
}

course-qa-agent-be.md

Course Q&A — Backend

A FastAPI service for a syllabus-grounded course assistant: ingest documents, retrieve relevant passages, and answer with citations instead of guesses.

Role

Team lead on a seven-person team. I worked across architecture, backend behavior, role-aware access, integration, testing, and delivery.

What the service does

• Authenticates students, trainers, and admins, then applies role-aware access to course material.
• Ingests course documents, splits them into retrievable chunks, and stores embeddings in Qdrant.
• Retrieves the most relevant context for a question and asks Gemini for an answer grounded in that material.
• Returns sources alongside the answer so the result can be checked, not just trusted.
• Exposes health, admin configuration, and test-connection endpoints for deployment checks.

Retrieval path

Stack

Engineering notes

The backend is a layered modular monolith: entrypoint and routers at the edge, schemas at the contract, services for business logic, and configuration kept out of code. Each epic owns a module, which keeps a large course project navigable.

Schema changes belong to Alembic. The container runs migrations on startup and fails clearly if it cannot, rather than silently creating tables. Local infrastructure uses the same ports as Docker so the two modes never fight.

course-qa-agent-fe.md

Course Q&A — Frontend

The Next.js application around the course assistant: a role-aware shell, a focused conversation surface, and citations that stay visible while reading an answer.

Role

Team lead on a seven-person team. I shaped the frontend foundation, the role-aware navigation, the chat experience, and the shared component layer used across the app.

What the app does

• Signs users in and routes students, trainers, and admins to the areas they are allowed to see.
• Provides a chat surface where a question, answer, and source citations stay in one readable view.
• Keeps course and document context close to the conversation instead of hiding it behind extra screens.
• Uses a shared set of UI primitives so new pages do not drift in spacing, color, or behavior.
• Includes admin settings for model and email configuration, plus a component styleguide for the team.

Request flow

Interface layers

Stack

Engineering notes

Design tokens live in one place so themes and spacing stay consistent. Navigation is configuration-driven: a route declares which roles may see it, and the shell handles the rest.

The client-side auth guard is treated as UX, not security. The backend still enforces every role and permission. That separation keeps the frontend helpful without pretending it is the source of truth.

devflow-platform.md

devflow-platform

A Jira/Linear-inspired project management platform for software teams, built as a long-term exercise in architecture, testing, and DevOps rather than a demo.

What it does

Teams can create workspaces, invite members with Owner/Admin/Member roles, manage projects, plan sprints, and move tasks across a Kanban board. The board updates in realtime through SignalR and supports assignee or epic swimlanes.

• Work management: projects, task keys, sprints, dependencies, blockers, time tracking, labels, and custom fields.
• Collaboration: comments, watchers, attachments, notifications, activity log, and a searchable knowledge base.
• Delivery signals: GitHub repository linking, PR tracking, webhooks with a retryable dead-letter queue, and task keys linked from commits.
• Insight: burndown charts, velocity metrics, and per-member cycle and lead-time reports.
• AI planning: a provider-agnostic agent that can plan work against project knowledge with self-approval rules.

Architecture

The backend follows Clean Architecture with CQRS and MediatR. Authorization is declarative, using pipeline behavior and workspace/project role checks so access rules do not scatter across handlers.

PostgreSQL stores the domain; Redis handles caching and realtime support; SignalR pushes board changes to connected clients. Authentication uses short-lived JWT access tokens with rotating refresh tokens, plus Google and GitHub OAuth.

Request path

Stack

What I focused on

The interesting problems were boundary problems: keeping a single active sprint per project, enforcing role rules in one place, modeling a dependency graph without making the board unreadable, and making webhook failures inspectable instead of lost.

The project is complete. It is public because the architecture and test suite are part of the point.

KCPM.md

KCPM

A waste-recycling platform that connects citizens, collectors, recycling enterprises, and administrators in one traceable workflow.

Context

Built for the Software Verification course as Team 36. The project was as much about professional test process as product behavior: requirements to Jira tasks, unit and end-to-end tests, API smoke tests, coverage, static analysis, and deployment checks.

My role

Team lead, CI/CD owner, and architect on a five-person team. I was responsible for the auth, collector, enterprise-collector, waste-category, SignalR, and role-based security modules. I also owned the xUnit and E2E coverage, SonarCloud quality gate, and the deployment pipeline.

What the platform does

• Citizen: create waste reports with GPS, photos, and category, then follow status and earn rewards points.
• Enterprise: manage collectors, assign collection jobs, and view reports.
• Collector: receive jobs and update collection status in the field.
• Admin: manage users, approve enterprises, and handle complaints.

Role flow

Stack

Quality work

The pipeline runs backend tests with coverage, frontend E2E, Postman smoke tests, SonarCloud analysis, and deployment. Allure reports publish to GitHub Pages, and CI results post back to Jira so failures stay visible where the work is planned.

skills.md

Skills

The tools I reach for most, and the projects where they actually did something useful.

Toolbelt

• PostmanAPI testing
• DockerContainers
• GitVersion control
• Claude CodeAgentic CLI
• Cheat EngineMemory
• x64dbgDebugging
• IDADisassembly
• SpotifyFocus

Capability stack

Languages

C#, Python, TypeScript, PHP, Kotlin, SQL, Luau.

Backend

ASP.NET Core, FastAPI, Laravel, Node.js. API design, authentication and authorization, relational data, background work, and modular monolith structure.

Frontend

React, Next.js, TypeScript, Tailwind CSS, shadcn/ui. App Router, role-aware shells, readable chat surfaces, and shared component systems.

AI and data

Gemini, LangChain, retrieval-augmented generation, Qdrant, PostgreSQL, MySQL, SQLite, Redis. Document ingestion, chunking, vector retrieval, source-grounded answers, and citations.

Infrastructure and practice

Docker, Docker Compose, GitHub Actions, RabbitMQ, SonarCloud, Allure, Jira. Clean Architecture, CQRS + MediatR, automated testing, code review, and Agile / Scrum delivery.

Where it shows up

Approach

Build the whole thing, then trace it back. Tests and documentation are part of the interface, not the cleanup after it.

contact.md

Contact

For work, questions, or a system worth taking apart. Email is the fastest route.

Channels

Good reasons to write

• A backend, AI, or full-stack problem that needs someone to read the whole system.
• A systems question, tool, or write-up worth comparing notes on.
• Source access to the private Course Q&A repositories.

Based in Ho Chi Minh, Vietnam · GMT+7

README.md

This is a workspace

Not a screenshot of an IDE. Everything here is part of the portfolio: open files from the explorer, search the workspace, switch themes, or type in the terminal.

What to open

• experience.md — the path to building AI products.
• projects.md — four systems and the decisions behind them.
• agent-skills.md — a catalog of agent skills with full SKILL.md detail and original sources.
• contact.md — open a channel.

Every content file is Markdown now. The stack details, repository names, and project boundaries are written out instead of hidden in sample code.

Shortcuts