Skip to content

Quoc Tran Trung - portfolio

I shipproduction systemsend to end.

Full-Stack Engineer · Technical Lead. AI cameras at airports, cloud platforms that scale, and tools that help small teams ship big things. Based in Hanoi.

Begin the storyOpen to lead roles

Chapter 01 · Origin

The first three years happened on a single laptop in Hanoi.

2021

Hanoi University

B.Sc. IT, Software Engineering - English programme. The PHP-on-WAMP era.

2022

Devpro J2EE certification

Java Web Full-Stack, eight months of writing servlets after class.

2023

HANU IT Youth Union - Logistics Lead

Ran logistics for the IT Department Freshmen Welcome and the Warm Spring volunteer campaign.

2024

British University Vietnam

PHP / Laravel internship in the ICT department. Built a barcode-detection pipeline that retired a manual data-entry workflow.

2024

HANU Youth Union - Secretary

Led the executive team across campus-wide events. First time owning a roadmap end-to-end.

Chapter 02 · Crossover · Mar - Jul 2025

Bitsness gave me my first AI automation gig.

The platform ran a YouTube channel's next video through five automated stages - research, ideation, scripting, storyboard, finishing - before a human editor ever opened it. My job was the backend pipeline that orchestrated it, the security around access, and the deploy pipeline that took it from laptop to staging.

  1. 01

    Video

    Source ingestion

  2. 02

    Analysis

    OpenAI + YouTube APIs

  3. 03

    Ideation

    Topic + angle generation

  4. 04

    Script

    Long-form narrative

  5. 05

    Storyboard

    Shot list + thumbnails

I led the backend split - modular NestJS services, repository pattern, role-based guards - and built the GitLab pipelines that took it from laptop to staging.

Chapter 03 · Lead arc · PathTech, Jul 2025 → Jul 2026

Six months in, I was leading the AI + cloud workstreams.

Joined as a Full-Stack Engineer, promoted to Technical Lead in Dec 2025, then seven months leading the AI and cloud workstreams. Each tile below is a real product running in production - tap any of them for the full story, the live screenshots, and the lessons learned.

Chapter 04 · AIBox closeup · two production deployments

Edge AI live at Noi Bai airport and a 110 kV substation.

Live at Petrolimex Aviation at Noi Bai International Airport and at EVN's 110 kV Mo Lao substation. One rugged box per site watches four cameras at once, spots safety violations the instant they happen, and pushes alerts to the cloud in under a second. No video ever leaves the site for AI - the inference runs right there on the device.

  • 0+

    YOLO models in prod

  • 0+

    REST endpoints

  • 0

    concurrent CCTV feeds

Models in the field

  • helmet detection
  • oil-spill
  • smoke + fire
  • license-plate recognition
  • face recognition
  • people control
  • petrolimex uniform

Streaming: WebRTC primary, RTSP for camera ingest, MJPEG fallback. WebSocket broadcasts every detection event for the cloud's live timeline.

Live frame from AIBox at Petrolimex Aviation, Noi Bai: a refueling truck and ground crew with YOLO bounding boxes drawn around helmets, vests, and a flagged PPE-negative person.
AIB-PA-NOIBAI-001 · LIVE1920×1080 · 25 fps
aivision.petrolimexaviation.comWebRTC ↔ Cloud

Chapter 05 · mnemov4.6.2 published

A memory layer that makes Claude Code remember what you've already taught it.

All the lessons you teach an AI coding assistant - your team's patterns, your design decisions, the gotchas you keep re-learning - collected into a small searchable graph on your own laptop. mnemo picks the most relevant pieces every time you ask the AI a question, and feeds them in automatically, with citations, capped at a small budget so it never floods the conversation. Open source, runs offline, shipping new releases continuously (the version above is live from GitHub).

Scoring breakdown

avg of top 3

  • 0.40Vector similarity
  • 0.15Typed-edge graph
  • 0.10Recency (90d half-life)
  • 0.10Memory type bias
  • 0.05Project scope
  • 0.20Lexical overlap
  • 100%

    top-1 accuracy

  • 1.000

    MRR

  • 17 ms

    median query

  • ≤ 800

    tokens injected per prompt

Top hits

  • 0.307

    mnemo-built

    mnemo (local-first knowledge memory for Claude Code) shipped through v1.1.0 - architecture, six-term scoring, plugin install flow.

    [project]
  • 0.217

    mnemo-release-workflow

    PR-only release workflow for mmct-jsc/mnemo: push to release/<ver>, open PR, merge in web UI - main is branch-protected.

    [feedback]
  • 0.185

    aibox-petrolimex

    Edge AI multi-camera surveillance on Jetson for Petrolimex Aviation - helmet / oil-spill / smoke / plate / face YOLO models.

    [project]

Live demo · the real UI

Open full-screen

The query box above is a six-node mock that mirrors the scoring; the embed is the real shipped demo. Source: github.com/mmct-jsc/mnemo.

Chapter 06 · Constellation

Tools I reach for, grouped by where they live in the stack.

Eight clusters, one stack. Languages feed Backends, Backends sit on Data, Data sits on Infra; AI/ML borrows from each.

Languages

  • Python
  • Go 1.25
  • TypeScript
  • Java
  • PHP
  • C#

Backend

  • FastAPI
  • NestJS
  • Spring Boot
  • Laravel
  • Express

Frontend

  • React
  • React Native
  • Vue 3
  • Tailwind

AI / ML

  • YOLO v8/v11
  • FaceNet / MTCNN
  • pgvector
  • Llama 3.1
  • OpenAI + Whisper
  • sentence-transformers

Data

  • PostgreSQL
  • MySQL
  • SQLite + sqlite-vec
  • MongoDB
  • Redis
  • Prisma
  • GORM

Infra

  • Docker
  • Kubernetes (GKE)
  • Nginx
  • HAProxy
  • MediaMTX RTSP
  • Terraform

Edge / IoT

  • NVIDIA Jetson (CUDA 12.8)
  • RTSP
  • WebRTC
  • WebSocket

Architecture

  • Clean Architecture
  • DDD
  • Microservices
  • Event-driven
  • Multi-tenant SaaS
  • Local-first

These are the typical seams the projects on this site cross. Reduced-motion users see the same layout without the draw-in.

Chapter 07 · Coda

Let's talk.

Open to Full-Stack Engineer and Technical Lead roles. Hanoi-based, GMT+7, response within one working day.