Works
Things I build to understand.
Projects, research experiments, and manuscripts across AI, learning systems, productivity, and energy.
Selected projects
T4 Speculative Decoding Acceleration
LLM speculative decoding acceleration framework for Qwen architectures using BitsAndBytes NF4/INT8 quantization and draft model distillation on NVIDIA T4.
LLM · Speculative Decoding · Qwen · PyTorchView projectGemma4 Inference Latency Benchmarks
Systematic quantization profiling, latency analysis, and Random Forest execution time prediction for Google Gemma on resource-constrained hardware.
LLM · Gemma · Quantization · BenchmarkingView projectNeural Compression CPU Benchmarks
Hardware-aware neural network compression suite featuring INT8 post-training quantization, structured L1 pruning, and low-rank tensor decomposition on CPU.
Compression · INT8 PTQ · SVD · CPU OptimizationView projectWARI Adaptive Residual Inference
Novel layer-granule residual CPU scheduling policy (WARI-L) optimizing working-set memory footprint during deep learning inference.
Inference · Memory Scheduling · SystemsView projectChinese-to-Vietnamese Transformer NMT
End-to-end Neural Machine Translation system engineered from scratch with SentencePiece subword tokenization, multi-head attention, and length-penalized beam search.
NLP · Transformer · SentencePiece · PyTorchView projectVOAI Olympic AI Comprehensive Curriculum
Master-level 38-notebook training curriculum (M0 to M7) covering PyTorch tensor engineering, algorithmic optimization, and competitive AI contest workflows.
Curriculum · PyTorch · Jupyter · AlgorithmsView projectSecure Binary Protocol Framework
Authenticated 6-byte packet framing specification with AES-256-GCM authenticated state machine and anti-replay counters for telemetry streaming.
Networking · Protocol Security · AES-GCM · SystemsView projectBoconic Book Sharing Platform
Distributed book-sharing and community platform built on FastAPI 0.110+, SQLAlchemy 2.0 Async, Telegram Bot aiogram 3.x, and Docker containers.
FastAPI · AsyncIO · Docker · PostgreSQLView projectSắc Việt Heritage Studio
Interactive 3D Vietnamese cultural heritage digital studio utilizing Next.js 15, React 19, Three.js, React Three Fiber, and SQLite WAL storage.
3D Web · Next.js 15 · Three.js · React 19View projectNeuralWork Knowledge OS
Local-first knowledge workspace featuring a TypeScript 5 monorepo, Node.js native SQLite WAL engine, and FTS5 BM25 search ranking.
TypeScript 5 · SQLite WAL · FTS5 · Local-firstView projectPrompTEA Prompt Optimization
Language model prompt optimization and systematic refinement framework leveraging NLP and iterative reward signals.
LLM · Prompt Engineering · NLPView projectNeural Network from Scratch
First-principles neural network with analytical backpropagation and finite-difference gradient verification (< 10^-6) on MNIST.
Deep Learning · First Principles · NumPy · CalculusView projectConservative Force Simulation
Scientific computing engine solving non-linear conservative potential ODEs via Runge-Kutta RK45 with CustomTkinter and MATLAB interfaces.
Physics · RK45 ODE · SymPy · SimulationView projectFocumia Productivity OS
Client-side deep work Pomodoro workstation with hierarchical task management, soundscape generators, and zero-server persistence.
Productivity · Vanilla JS · Local StorageView projectResearch papers
LogicPT–GT
Logic · problem solving · structured thinking · Vietnamese
AIAPE–Education
AI · programming education · learning systems · English
Peer-reviewed work will be added after formal release, with venue, DOI, abstract, and citation metadata.
Build with me
I am open to 2026 hackathons and student collaborations in AI, education, data, and sustainability.