Work & Research
Featured project
Benchmarking Streaming ASR for Vietnamese
Publication — Proceedings | 6th International Conference on Computing and Communication Networks (ICCCNet 2026)
Proceedings paper, “Benchmarking Streaming ASR for Real-time Deployment: A Robustness Scorecard and Error Taxonomy for Vietnamese,” presented at the 6th International Conference on Computing and Communication Networks (ICCCNet 2026). The conference was organized by Manchester Metropolitan University in Manchester, United Kingdom, from 17–19 July 2026.
My work
Built the manifest-first data standardization and ASR evaluation workflow Implemented result aggregation, plots, bootstrap confidence intervals, and report-ready artifact generation
Outcome
- •Implemented a standardized JSONL manifest and strict Vietnamese normalization flow for fair ASR evaluation
- •Benchmarked models across offline and pseudo-streaming modes with WER/CER, real-time factor, latency proxy, and stability metrics
Featured project
mondaiy — AI Work Navigator
VinUni AI in Action Program, Cohort 3 | Evidence-backed personal work assistant
Built and led delivery of mondaiy (P-004) in VinUni's AI in Action Program, Cohort 3. The AI work navigator brings together user-permitted messages, email, calendar events, tasks, and long-term memory. It creates evidence-backed daily briefings, answers questions with source links, extracts actionable work, and prepares actions for user approval.
My work
Owned product definition, roadmap, source-of-truth documentation, and delivery coordination across a four-person team Authored the PRD, user stories, acceptance criteria, and product safety boundaries for provenance, source conflict handling, and approval-gated actions Integrated and reviewed vertical slices spanning FastAPI, LangGraph, React, browser extension, provider connectors, and PostgreSQL/pgvector retrieval Owned test coverage, CI/CD, production monitoring, latency and cost baselines, and deployment ownership for the Mondaiy delivery path
Outcome
- •Delivered an active MVP/pilot product that consolidates permitted Gmail, Google Calendar, Discord, Jira, and Outlook sources
- •Implemented evidence-backed daily briefs, source-linked Q&A, actionable-work extraction, and user-approved task or calendar actions
Featured project
GreenFlow: Agentic Digital Twin for Building Energy
Finalist, Vin Green Future Hackathon 2026
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.
My work
Designed the LangGraph orchestration flow across semantic analysis, prediction, control, simulation, policy, approval, and response composition Built function-calling tools that query structured building telemetry and KPI data through parameterized SQL instead of LLM-generated SQL Developed graph-grounded and hybrid RAG retrieval for building entities, equipment relationships, policies, reports, and historical context Implemented the Ask GreenFlow chatbot with tool traces, cited sources, linked building entities, conversation persistence, and configurable LLM providers
Outcome
- •Selected as a finalist at the Vin Green Future Hackathon 2026
- •Delivered an agentic workflow that grounds recommendations in building state, runs counterfactual simulation, applies policy guardrails, and routes risky actions for human approval
Featured project
Minute – Retrieval-Based AI Meeting Co-Host for BFSI
VNPT AI Hackathon 2025 | Desktop/Web meeting workflow with citations and audit trails
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.
My work
Worked alongside one other code contributor on the Minute implementation Designed the pgvector data model, document-ingestion flow, and metadata/ACL filtering for permission-aware retrieval Contributed to LangGraph routing, FastAPI APIs, and the real-time WebSocket meeting pipeline
Outcome
- •Built end-to-end AI meeting workflow: Pre-meeting (agenda + pre-read) → In-meeting (live transcript, recap, ADR extraction) → Post-meeting (executive summary, MoM, task sync)
- •Contributed to stage-aware LangGraph routing for the Pre/In/Post meeting workflow
AI Tour Guide Generator – 3D Building Tour Builder
Full-stack AI tour builder with panorama viewing, LLM image narration, TTS, and Supabase persistence
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.
My work
Built the tour builder UX, FastAPI endpoints, LLM provider abstraction, TTS endpoint, and persistence flow Deployed the Next.js frontend and added demo fallbacks for local or no-key usage
Outcome
- •Shipped a deployed frontend for building and playing AI-generated building tours with panorama and normal-image support
- •Implemented multi-provider LLM runtime for image description and tour-script generation with Gemini, Groq, OpenAI, Claude, and mock fallback
HPC MLOps Object Detection with YOLO Knowledge Distillation
Real-time traffic object detection deployment with Docker, FastAPI, Gradio, ONNX, and TensorRT
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.
My work
Worked on the MLOps project structure, model artifact workflow, serving pipeline, and benchmark interpretation Packaged FastAPI/Gradio serving, Docker deployment path, and MLflow/MinIO tracking evidence
Outcome
- •Built a teacher-student YOLO training workflow with Knowledge Distillation for traffic object detection
- •Student KD reached mAP50 0.725 and mAP50-95 0.490 while reducing parameters from 58.82M to 2.51M and model size from 112.85 MB to 5.14 MB
Edge-Cloud Reinforcement Learning for HVAC Control Across Vietnamese Climate Zones
Transferable HVAC control research with HOT building archetypes and Vietnam weather contexts
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.
My work
Prepared the reproducible experiment structure, summary metrics, and manuscript figures Packaged result tables, trained policy artifacts, cloud run notes, and the final report workflow
Outcome
- •Built reproducible experiment and manuscript artifacts for transferable HVAC control across Vietnamese building/weather contexts
- •Compared controller families using energy, comfort violation rate, temperature deviation, action instability, deployment score, and transfer regret metrics
Customer Segmentation with Excel, Orange, and K-Means
Retail customer clustering from Online Retail transactions
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.
My work
Led the portfolio-safe artifact: data inspection, EDA framing, and customer-level feature design Built and interpreted the K-Means segmentation workflow and final report
Outcome
- •Processed 541,909 raw transaction rows into 349,203 valid records after CustomerID filtering, transaction cleaning, UK-only filtering, duplicate handling, and TotalPrice creation
- •Engineered 16 customer-level behavioral features covering recency, frequency, monetary value, product diversity, invoice behavior, and basket characteristics
KKBox Real-time Customer Churn BI Dashboard
UEH Business Intelligence course project with streaming analytics and decision support
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.
My work
Worked on the feature-store-first BI workflow and dashboard productization Integrated batch/streaming artifacts into FastAPI endpoints and the React decision-support dashboard
Outcome
- •Delivered a 3-tab decision-support dashboard spanning descriptive analysis, predictive-proxy scoring, and prescriptive scenario simulation
- •Built end-to-end near real-time data flow: replay logs -> Kafka -> Spark Structured Streaming -> ClickHouse -> FastAPI -> React dashboard
TomatoHub – AI-Powered Relief Campaign Platform
LotusHacks x HackHarvard x GenAI Fund Vietnam Hackathon submission
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.
My work
Implemented role-based campaign and supporter workflows in the full-stack monorepo Wired AI-assisted campaign drafting, prioritization, and recommendation features into the product flow
Outcome
- •Shipped a monorepo product with public pages, role-based dashboards, campaign lifecycle management, donation flow, and volunteer registration flow
- •Implemented QR-based volunteer and goods checkpoint logic alongside public transparency logs for auditability
Vietnamese Medical Information Extraction (NER + Relation Extraction)
UEH NLP course final project with semi-supervised IE pipeline for medical text
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.
My work
Worked on the demo interface, BERT-based NER workflow, silver-data generation, and vectorization functions Implemented evaluation logic for comparing relation extraction model variants
Outcome
- •Hybrid semi-supervised RE achieved 81.25% accuracy and 0.631 Macro-F1 (MLP + BERT)
- •Semi-supervised approach improved F1 from 0.599 (Standard) to 0.631 (Hybrid) with silver data augmentation
VN Stock Analytics – Investment Decision Support System
Multi-model data mining with LLMs reasoning for Vietnamese banking stocks
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.
My work
Designed the system architecture, data pipeline, feature set, and walk-forward validation workflow Trained XGBoost models for return, risk, direction, and regime prediction, then connected results to the app layer
Outcome
- •Return Regression achieved MAE 0.094 and RMSE 0.119 on 21-day log-return prediction
- •Risk Model achieved 0.98 correlation between predicted and actual volatility
Vietnam Weather Prediction with Softmax Regression
Multiclass weather classification using time-series feature engineering
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.
My work
Collected Open-Meteo data for 34 provinces and built lag, cyclic, difference, and accumulation features Trained the Softmax Regression model, tuned hyperparameters, and deployed the Streamlit demo
Outcome
- •Multiclass classification achieved 65.6% accuracy with macro F1-score 0.644
- •Feature engineering improved accuracy from 63.9% to 65.6% (+1.7pp)
Vietnamese Fake News Detection: Deep Learning vs Transfer Learning vs LLMs
First Prize @ UEH BIT Faculty Research | Presented at NCTD 2025 National Conference
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).
My work
Built the evaluation framework for BiLSTM, PhoBERT fine-tuning, and LLM prompting baselines Handled EDA, class-imbalance strategy, model comparison, and efficiency-performance analysis
Outcome
- •First Prize in Faculty-level Research Competition at Business Information Technology (BIT) Department, UEH
- •Paper presented at National Conference on Technology and Design 2025 (NCTD 2025) – Shaping Vietnam's Digital Future
Breast Cancer Ultrasound CAD: Sequential vs Multi-task Deep Learning
BIT Genesis Research Award 2025 @ UEH | Presented at NCTD 2025 National Conference
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).
My work
Preprocessed BUSI ultrasound data and implemented U-Net/EfficientNet model variants Ran ablation experiments, statistical tests, and contributed to the research paper
Outcome
- •BIT Genesis Research Award 2025 at Business Information Technology Department, UEH
- •Paper presented at National Conference on Technology and Design 2025 (NCTD 2025) – Shaping Vietnam's Digital Future
GA Maximum Flow Solver – Network Optimization with Genetic Algorithm
Interactive visualization of evolutionary approach for Maximum Network Flow Problem
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.
My work
Led a 4-member team and owned the PyQt5 graph editor, parameter-tuning UI, and GA visualization flow Integrated the GA solver with comparison views and prepared the final report
Outcome
- •Achieved up to 100% optimality ratio on graphs with ≤30 nodes, competitive with Ford-Fulkerson exact solution
- •Path-based crossover maintains flow conservation constraint, avoiding invalid offspring after genetic operations
Top 3 – Humanitarian Logistics Hackathon
Smart surplus-food allocation for underserved communities
Collaborated with a cross-university team to build a logistics solution combining data management, ML allocation, and IoT warehouse tracking to reduce food waste.
My work
Worked on the allocation logic for matching surplus-food supply with community demand Designed the core data structure and helped develop the MVP workflow for the hackathon demo
Outcome
- •Top 3 finalist across HCMC universities
- •Proposed ML-driven allocation reducing surplus mismatch