quang/ai
VI Hire me

Open to AI Engineer roles · Ho Chi Minh City / Remote

Trần Ngọc Quang I train vision models,
make them fast, and ship them.

AI Engineer across computer vision, 3D Gaussian Splatting, TensorRT deployment and LLM systems — from multi-day pre-training runs to engines that hit their latency budget on real hardware.

See the work Resume
  • Top 4 of 1,600+ teams · Viettel AI Race 2026, BTS digital twin
  • 45→12ms per frame · YOLOv8-Pose, FP16 TensorRT on RTX 3060
  • 2→6 live RTSP streams per GPU in production
  • 98.9% plate-level accuracy · Vietnamese LPR benchmark

01 About

The unglamorous half of AI —
taking a model from a paper to production.

I'm an AI Engineer in Ho Chi Minh City and a B.Sc. in Artificial Intelligence from FPT University (GPA 3.53/4.0, Top 5 Outstanding IT Students, Spring 2026). Most of my time goes into the part after the paper: multi-day pre-training runs, TensorRT engines that refuse to build, and CUDA kernels that overflow once the images get big enough.

Lately that has meant 3D Gaussian Splatting for Viettel AI Race 2026 (Top 4 with team Zillexa), a real-time patient-monitoring system running on live hospital cameras at Whammy Tech, inference-efficient ConvMAE backbones for my capstone, and LLM products with my team at TSC.

Tiếng Việt Trần Ngọc Quang — Kỹ sư AI (AI Engineer) tại TP. Hồ Chí Minh, chuyên thị giác máy tính (Computer Vision), tái dựng 3D bằng Gaussian Splatting, tối ưu mô hình với TensorRT/ONNX và xây dựng hệ thống LLM. Top 4 Viettel AI Race 2026 (bảng BTS Digital Twin), cử nhân Trí tuệ nhân tạo Đại học FPT. Đang tìm vị trí AI Engineer / Computer Vision Engineer.

  • Train vision models

    Self-supervised / MAE pre-training at ImageNet scale, multi-GPU DDP and mixed precision — with ablations designed so the comparison is actually fair.

  • Make them fast

    PyTorch → ONNX → TensorRT, FP16 profiling, C++ post-processing. Latency, throughput and VRAM measured per pipeline, not guessed from parameter counts.

  • Rebuild the world in 3D

    Gaussian Splatting with gsplat — sharded training across GPUs, ensemble fusion, camera-model plumbing, and patching the rasterizer when it breaks.

  • Build LLM systems

    LLM backends with live CRM context, vision-LLM content pipelines, retries, circuit breakers, rate limits — and the Postgres, Redis and Docker they live on.

02 Selected work

Models that left the notebook.

Top 4 · 1 of 12 finalists 3D VisionTSC · Team Zillexa

BTS Digital Twin — Viettel AI Race 2026

Novel view synthesis of telecom towers from drone captures with 3D Gaussian Splatting, then large urban scenes under an inference-time limit on an NVIDIA H200 in the final.

  • gsplat MCMC backbone; SIMPLE_RADIAL cameras handled by undistort → pinhole training → re-distort
  • Multi-backbone ensembles fused with SELECT-median; training sharded across 4×A5000
  • Final round: capped MCMC, weight decay, MVGS multi-view gradient accumulation, fine-tuned SCUNet restoration, count-based pruning for render time
  • Audited gsplat's CUDA projection and patched an int32 intersection overflow at full resolution; Hopper (sm_90) Docker build
  • gsplat
  • CUDA
  • PyTorch
  • Docker
  • H200
Case study Private · walkthrough on request
VisionWhammy Tech · production

iPBMS — real-time patient monitoring

Fall and seizure detection on live RTSP hospital streams, running on-premise on two production cameras.

fall-detection accuracy
>85%
fall → alert
<2 s
ms / frame
45→12
streams / RTX 3060
2→6
  • YOLOv8-Pose + MoveNet Thunder, joint-angle analysis over 17-keypoint tracking
  • ByteTrack IDs, per-track state machine and hip-velocity threshold to cut false alarms; Normal / Warning / Danger over WebSocket + Supabase Realtime
  • Incident reports: keyframe selection → captioning → translation, delivered via FCM
  • TensorRT FP16
  • C++ NMS
  • ByteTrack
  • Supabase
Case study
VisionCapstone GSU26AI07

Inference-Efficient ConvMAE

Four hybrid backbones — Ghost, Transformer, Mamba-2, Bi-Mamba — MAE-pre-trained on ImageNet-1K and benchmarked under one bias-controlled ONNX → TensorRT protocol. The Ghost arm came out faster and lighter than ConvMAE-Base at comparable face-ID accuracy.

  • DDP
  • Mamba
  • TensorRT
Case study Repo
VisionLive demo

GhostConvMAE Face Demo

A live PyTorch ⇄ TensorRT switch on a 4 GB GTX 1650 with per-forward latency and peak-VRAM accounting. The self-test harness passes 30/30 task × model × backend combos — and caught a CASIA RecordIO label-corruption bug.

  • Gradio
  • ONNX
  • TensorRT
Demo Repo
VisionPaper

Face-Aware MAE

Landmark-guided masking steers MAE pre-training toward eyes, nose and mouth. A controlled study on CelebA cut identity-critical reconstruction error by 17.5% versus block-wise masking — with zero added parameters.

  • Self-supervised
  • ViT
Case study Repo
VisionBenchmark

Vietnamese License Plate Recognition

YOLOv5 / v8 / v11 × three OCR pipelines on real RTSP traffic data. Best pair (YOLOv11 + CNN-BiLSTM-CTC) reaches 98.9% plate-level accuracy, 0.994 mAP@0.5, and 91% on 15,000 night-time frames.

  • YOLOv11
  • CTC
  • OpenCV
Case study Repo
LLMTSC · client work

AI consulting console

LLM advisor bots for an education provider with CRM context injected every turn, a fire-and-forget pipeline caching suggestions and lead summaries in Redis, a persona/prompt compiler, output sanitizer, scoped API keys, rate limiting, retries and a circuit breaker.

  • TypeScript
  • PostgreSQL
  • Redis
Private
LLMIn progress

Vision-LLM short-video pipeline

ETag-incremental MinIO indexing, ffmpeg frame sampling, vision-LLM classification against a strict JSON schema into SQLite FTS5, and a facts corpus for grounded scripts — tested with moto and httpx MockTransport.

  • Python
  • MinIO
  • FTS5
Private
SystemsTSC

License control plane

Ed25519-signed licenses with a signed revocation list, an offline grace period, a CLI, and a JWT-protected admin UI.

  • Node.js
  • Ed25519
  • JWT
Private
VisionEdge

Vision Edge Healthcare System

Open-source edge release of fall/seizure monitoring: multi-camera RTSP, YOLOv8-Pose + MoveNet, Supabase Realtime + FCM alerts, under 3 s from detection to notification.

  • Docker
  • RTSP
  • FCM
GenAI

GenBGM API

A Dockerized GPU service that generates background music from chat context with MusicGen, tuned to fit in 4 GB of VRAM.

  • MusicGen
  • Docker
Repo
LLM

AI Talent Bot

A Telegram bot that parses PDF/DOCX CVs and scores candidates against a job description with Gemini.

  • Gemini
  • Telegram
Repo

03 Experience

Where the work shipped.

  1. Jun 2026 — Present

    AI Engineer @ TSC · Team Solo Code

    Ho Chi Minh City · three-person core team

    Client AI software and AI challenges: the Viettel AI Race 2026 BTS digital twin (Top 4), an LLM consulting console for an education provider, a vision-LLM short-video pipeline and a signed-license control plane.

  2. Aug 2025 — Present

    AI Engineer, Computer Vision @ Whammy Tech

    Ho Chi Minh City · contract, core team · in parallel with TSC

    Built the iPBMS patient-monitoring pipeline end to end: pose-based fall detection on live RTSP, FP16 TensorRT with C++ post-processing (45 → 12 ms/frame, 2 → 6 streams per RTX 3060), false-alarm control and multilingual incident reports. Operates on-premise on two production cameras with weekly stakeholder reviews.

  3. May 2025 — Aug 2025

    Automation Engineer Intern @ FPT Software Philippines

    Manila · OJT

    RPA + Arduino attendance check-in prototype for 120+ employees, built on four reusable UiPath components.

  4. Jun 2023 — Dec 2024

    IT Supporter, Technical Support @ FPT University

    Ho Chi Minh City

    Exam-system operations across two academic years: imaging, network checks, live login/CRM fixes and root-cause ticketing.

  5. Sep 2022 — Sep 2026

    B.Sc. Artificial Intelligence @ FPT University

    GPA 3.53 / 4.0

    Capstone GSU26AI07: Inference-Efficient ConvMAE for Universal Visual Recognition Tasks. Top 5 Outstanding IT Students (Spring 2026) · Top 100 (Fall 2025).

04 Stack

One pipeline, end to end.

The tools I reach for at each stage of getting a model into production.

  1. 01

    Train

    • PyTorch
    • TensorFlow
    • timm
    • HF Transformers
    • MAE / SSL
    • ViT
    • Mamba
    • DDP
    • Mixed precision
  2. 02

    Perceive

    • OpenCV
    • YOLO
    • Pose estimation
    • ByteTrack
    • OCR
    • 3DGS · gsplat
    • RTSP multi-stream
  3. 03

    Optimize

    • TensorRT
    • ONNX
    • CUDA
    • C++
    • FP16
    • Profiling
  4. 04

    Reason

    • LLM backends
    • Vision-LLM
    • Prompt / persona pipelines
    • Gemini API
    • AI agents
  5. 05

    Serve

    • Python
    • TypeScript
    • Node / Express
    • React
    • PostgreSQL
    • Redis
    • MinIO / S3
    • Docker
    • WebSocket
    • Supabase

05 Honors & credentials

Competitive since the Olympiad days.

  • 2026

    Prospective Prize — Top 4 of 1,600+ teams

    Viettel AI Race 2026 · Track 1: BTS Digital Twin

  • 2026

    Top 5 Outstanding IT Students

    FPT University · Spring 2026

  • 2025

    Top 100 Outstanding IT Students

    FPT University · Fall 2025

  • 2021

    Consolation Prize

    Vietnam National Olympiad in Informatics (VOI)

  • 2020

    Silver Medal

    Central Coast & Northern Coast Programming Contests

Certifications

  • DeepLearning.AI — TensorFlow Developer Professional Certificate
  • IBM — AI Enterprise Workflow Specialization
  • DeepLearning.AI — NLP with Attention Models · Sequence Models
  • UC San Diego — Machine Learning With Big Data
  • IBM — Containers: Docker, Kubernetes & OpenShift
  • IBM — Microservices & Serverless

Languages

  • Vietnamesenative
  • EnglishEF SET 72/100 · C2
  • Japaneseelementary

06 Live from GitHub

Recently pushed.

github.com/QuangCler

07 Contact

Have a model that needs
to run faster?

Open to AI Engineer roles in computer vision, 3D or LLM systems — Ho Chi Minh City or remote. The quickest way to reach me is email.

Send a message

POST /api/contact → D1

A role, a project or a model that's too slow — anything works.

esc