Turns a one-line description of a research field into a map of its literature: topics, trends and gaps. An agent answers questions about the field, citing only papers it actually retrieved.
A tax-compliance assistant for Vietnamese online sellers. It answers legal questions with a citation on every claim and reconciles sales against declared revenue. Built by a team of four that Khai led.
A realtime two-player board game built as a platform. It has a deterministic rules core, one set of typed schemas that validates, documents and tests the API, and a separate tournament service for bots.
An online shop and workshop booking site for a Hanoi pottery studio, built on WordPress with Pancake POS as the system of record. Its public demo replays recorded POS traffic.
Booking, a self-assessment and a chat assistant for a clinic that treats period pain with traditional medicine. The AI stays strictly inside clinician-approved content.
A course built as software. A Python pipeline generates the course material and executes it as a test, and a Next.js platform runs learners' code either in the browser or on a pool of remote GPUs.
A computer-vision system split across a network. A stateful client handles the camera and the dashboard, and a stateless FastAPI service runs YOLOv8 on a GPU.
Built a 12.5M-pair dataset for Vietnamese diacritic restoration and implemented RNN, GRU, LSTM and Transformer seq2seq models from scratch in PyTorch.
Optimising GPU inference for ResNet50 and DistilBERT with FP16, INT8, TorchScript/torch.compile and ONNX.
A real-time license plate recognition system: a fine-tuned YOLOv5 model with OCR, a FastAPI and PostgreSQL backend and a React frontend, deployed with Docker.
A quantitative trading system built from scratch in PyTorch: a Transformer that predicts BUY/SELL/HOLD signals from financial indicators and historical trading data.
Fraud detection on the Kaggle credit card dataset, where 0.172% of transactions are fraud, comparing four classifiers with SMOTE and ADASYN resampling.
Sentiment classification of 1.6M tweets (Sentiment140) with Naive Bayes and SVM, run on Apache Hadoop and Apache Spark.
YOLOv1 rebuilt from scratch in PyTorch, with its own dataset class and loss function, and CBAM attention added to the feature extractor.
