Built a reproducible benchmark for Vietnamese ASR models across offline and simulated streaming settings. The work standardizes manifest generation, Vietnamese text normalization, WER/CER evaluation with bootstrap confidence intervals, latency proxy measurement, real-time factor reporting, and publication-ready result aggregation. The paper was presented at ICCCNet 2026 in Manchester, United Kingdom.

Built the AI layer and web experience for GreenFlow, a simulation-first digital twin that helps building operators investigate energy use and evaluate safer control actions. The system combines LangGraph agents, function-calling over operational data, graph-grounded RAG, and an operator chatbot with simulation, policy gates, human approval, and auditable execution traces.

Minute standardizes the meeting lifecycle for BFSI/LPBank enterprises: pre-meeting context gathering, real-time in-meeting assistance, and post-meeting minutes + action items generation, with citations, audit trails, and access control. The system uses stage-aware routing, a real-time WebSocket pipeline, permission-aware retrieval with pgvector, and human confirmation for tool actions.

Built and led delivery of mondaiy (P-004), an AI work navigator that brings together user-permitted messages, email, calendar events, tasks, and long-term memory. The product creates evidence-backed daily briefings, answers questions with source links, extracts actionable work, and prepares actions for user approval.

Built a deployed full-stack application for creating AI-guided 3D building tours. The system lets users create tour projects, upload room photos or panorama images, select LLM providers, generate scene descriptions, edit guide scripts, render Vietnamese female narration, and open an immersive fullscreen tour player with audio and floating room information.

Built an HPC-oriented MLOps pipeline for real-time traffic object detection. The project trains YOLO teacher, student baseline, and student knowledge-distilled models, compares accuracy-latency trade-offs, exports optimized artifacts, serves inference with FastAPI and Gradio, and wraps the system with Docker Compose, MLflow, MinIO, and monitoring components.

Research and reproducibility package for transferable HVAC control in Vietnamese climate zones. The project compares static, ASHRAE-style, cloud-only, edge-only, and edge-cloud adaptive controllers using HOT building archetypes, EnergyPlus-oriented simulation artifacts, transfer metrics, deployment scores, comfort violations, energy consumption, and latency analysis.

Built a customer segmentation workflow for Online Retail transaction data using Excel for raw-data inspection and Orange for visual preprocessing, feature engineering, clustering, and interpretation. The portfolio report version has been regenerated under Thai Hoai An only, with all other names removed from the public artifact.

Built a near real-time Business Intelligence system for KKBox churn monitoring and retention decision support. The project combines Kafka log replay, Spark Structured Streaming, ClickHouse OLAP storage, FastAPI APIs, and a React dashboard to deliver descriptive, predictive-proxy, and prescriptive analysis in one workflow.

TomatoHub is a full-stack platform for charity operations that helps organizations launch campaigns faster, lets supporters donate or volunteer with clearer trust signals, and keeps campaign activity transparent. The product combines role-based workflows, QR-based check-in/check-out, public transparency logs, and AI-assisted campaign drafting plus supporter recommendations.

UEH NLP course final project building an end-to-end Information Extraction system for Vietnamese medical text. Implemented a pipeline architecture (NER → Entity Pairing → Relation Extraction) inspired by PURE, recognized 5 entity types and 4 relation types, and used semi-supervised hybrid learning with silver data to overcome limited labeled data.

Data Mining course final project building comprehensive analytics system for 14 Vietnamese banking stocks in VN30 index. Integrated multi-source data (market OHLCV, financial reports, macro indicators, news sentiment) and developed 4 XGBoost models: Return Regression, Direction Classification, Risk Forecasting, and Regime Detection. LLMs layer provides reasoning and investment recommendations in natural language.

Data Visualization course final project building weather prediction model for 34 Vietnamese provinces using Softmax Regression. Collected 265K+ records from Open-Meteo API (2005-2025), engineered lag features, cyclic encoding for seasonality, and accumulation features. Classified weather into 3 groups (Clear/Cloudy, Drizzle, Rain) with rigorous train/test split by time.

Faculty-level research project providing comprehensive comparative evaluation of machine learning approaches for Vietnamese fake news detection. Systematically analyzed three major model families: traditional deep learning (BiLSTM with Word2Vec/FastText), transfer learning (PhoBERT frozen/fine-tuned), and large language models (Qwen2.5-7B, Llama-2-7B, DeepSeek) across zero-shot and few-shot paradigms. Evaluated on ReINTEL dataset (9,713 Vietnamese social media posts with 83.2% real vs 16.8% fake class imbalance).

Faculty-level research comparing Sequential and Multi-task Learning architectures for breast cancer diagnosis from ultrasound images. Built on U-Net with EfficientNet-B4 backbone, systematically evaluated Deformable Convolution and Capsule Network modules through ablation study. Evaluated on BUSI dataset (780 images: Normal/Benign/Malignant) with rigorous statistical testing (Shapiro-Wilk, Mann-Whitney U, Kruskal-Wallis, Tukey HSD).

Artificial Intelligence course final project applying Genetic Algorithm to solve Maximum Network Flow Problem. Implemented custom GA operators: path-based crossover to maintain flow conservation, adaptive mutation for escaping local optima, and balance flow mechanism. Built interactive Python GUI for real-time graph editing, parameter tuning, and algorithm comparison with Ford-Fulkerson.

Collaborated with a cross-university team to build a logistics solution combining data management, ML allocation, and IoT warehouse tracking to reduce food waste.