Nguyễn Đình Khải

Based in Vietnam, I am an AI researcher and AI engineer with a strong passion for artificial intelligence and data-driven technologies. My academic background in computer science and hands-on experience with projects in machine learning, deep learning, natural language processing, and financial technology have equipped me with both theoretical knowledge and practical skills. I have worked on projects ranging from image recognition and fraud detection to building intelligent chatbots and real-time systems. This journey has given me valuable expertise in data analysis, programming, and problem-solving. I am deeply motivated to create AI solutions that not only advance technology but also deliver real-world impact. Innovation, continuous learning, and pushing the boundaries of what AI can achieve are at the core of my work, and I aspire to build impactful projects that shape the future of intelligent systems.

Machine Learning

PyTorch · Transformer · Seq2Seq · YOLOv5 / YOLOv1 · CBAM · XGBoost · SMOTE / ADASYN · scikit-learn

NLP

Vietnamese NLP · Diacritic restoration · BLEU / ChrF++ · Sentiment classification

Systems & Deployment

FastAPI · Docker · PostgreSQL · React · Apache Spark · Apache Hadoop

Inference & Performance

TensorRT · ONNX Runtime · TorchScript · torch.compile · FP16 / INT8 · CUDA profiling

• Detoura2026

  A TypeScript monorepo that turned a single online board game into a platform. Rules plug into one typed interface, the server treats a match as opaque state, one set of Zod schemas is both the validator and the published API spec, and a second service runs automated tournaments between bots over WebSocket.

  Workspace packages 9 · Interfaces under contract 5 · Protocol messages 14

• AI Pipeline Mastery2026

  A teaching system built as software: a Python pipeline that generates course artifacts from source and verifies them by executing them, and a Next.js platform that authenticates learners, serves material from private storage, and runs their code either in the browser or on a pool of remote GPUs.

  Slides generated 383 · Figures rendered 129 · Execution lanes 2

• Classroom Trash Detection2026

  A computer-vision system split across a network boundary. A stateful client owns capture, presentation and storage on ordinary hardware; a stateless FastAPI service owns the model on a GPU; and everything between them — bandwidth, latency, failure and portability — is engineered at the seam.

  Model formats 3 · Upload width 640 px · Deploy targets 3

• Neural Machine Translation2025

  A 12.5M-pair Vietnamese diacritic-restoration corpus, and four seq2seq architectures built from scratch in PyTorch to translate undiacritised text back into correct Vietnamese.

  BLEU 81.31 · ChrF++ 88.57 · Training pairs 12.5M

• Optimization for AI Inference Engines on GPUs2025

  A systematic study of GPU inference optimisation across ResNet50 and DistilBERT — FP16, INT8 quantisation, TorchScript, torch.compile and ONNX — measured rather than assumed.

  FP16 throughput 4× · INT8 speedup 2–5× · Graph optimisation 1.2–1.8×

• License Plate Recognition System2025

  A fine-tuned YOLOv5 detector coupled with OCR, wrapped in a FastAPI service and a React frontend so citizens can look up their own traffic violations. Runs at ~12 FPS.

  Throughput ~12 FPS · Pipeline stages 4

• Quantitative Trading System2025

  A quantitative trading system built from scratch in PyTorch — a Transformer classifying BUY/SELL/HOLD from technical indicators and Time2Vec embeddings over 3.4M training samples.

  Validation accuracy 92.75% · Test accuracy 92.50% · Training samples 3.4M

• Credit Card Fraud Detection2024

  Fraud detection on a dataset that is 0.172% positive — where accuracy is a useless metric and the whole problem is resampling strategy and the recall/precision trade-off.

  ROC-AUC 0.9817 · Recall 0.816 · Positive rate 0.172%

• Large-Scale Text Sentiment Classification2024

  Sentiment classification on Sentiment140 with Naive Bayes and SVM, run on both Apache Hadoop and Apache Spark to measure what the execution engine actually costs you.

  Dataset 1.6M · SVM accuracy 80–82% · Naive Bayes accuracy 75%

• YOLOv1 With CBAM From Scratch2024

  A ground-up PyTorch reimplementation of YOLOv1 — custom dataset, custom loss — extended with CBAM attention to measure what attention contributes to a one-stage detector.

  Train mAP 0.8829 · Best val mAP 0.6994 · Parameters 269.2M

• Third Prize, Student Scientific Research Award — Artificial Intelligence — VNU University of Engineering and Technology2026
• Third Prize, Student Track — Humanitarian Logistics Hackathon — Humanitarian Logistics Hackathon2026
• Academic Encouragement Scholarship — University of Engineering and Technology2024
• Second Prize, Provincial Excellent Student Contest in Physics — Thai Nguyen Province2023
• Third Prize, Provincial Excellent Student Contest in Informatics — Thai Nguyen Province2021