I build generative AI systems and secure cloud infrastructure — from multilingual medical conversation summarization (Qwen3-4B LoRA) to LLM agents that find and patch Infrastructure-as-Code vulnerabilities.
Information Technology student at VNUHCM University of Science, working at the intersection of AI Agent & LLM systems and security engineering. My work spans multilingual document summarization pipelines across 17 languages, static-analysis tooling powered by fine-tuned code LLMs, and scalable cloud backends with automated CI/CD.
Most of my work starts with the same question: what breaks if this model, policy or pipeline is wrong? The answer shapes the architecture — leak-free evaluation splits, AST-grounded context for LLM findings, ABAC at the database layer, and multi-layer validation gates that elevate response constraint scores to 99.5%.
Multi-layer constraint gates and access control treated as build requirements, not afterthoughts.
Containerised services and CI/CD pipelines that deploy with zero downtime.
AST parsing, policy analysis and network fundamentals behind every AI feature.
Latency, precision and false-positive rates tracked before and after every change.
The stack I reach for when a system has to be accurate, fast and safe at the same time.
Fine-tuning, agent orchestration and evaluating models that ship to production.
Policy-as-code, AST static analysis and data-protection engineering.
Reproducible builds and zero-downtime delivery pipelines.
APIs, vector search and the distributed interfaces that glue services together.
Four engineering tracks across multilingual medical LLM summarization, AI-assisted infrastructure security, agentic cloud backends and applied machine learning. Explore interactive walkthroughs and architectural dataflows below.
Research Contributor · Research Lab
End-to-end intelligent clinical dialogue summarization pipeline across 17 languages using fine-tuned Qwen3-4B LoRA and mT5.
Pre-computed sample outputs from the project · no model runs in your browser
Fine-tuned Qwen3-4B LoRA & 5-Layer Constraint Verification Gate
Tái khám tăng huyết áp định kỳ, phù mắt cá chân hai bên, hồi hộp thoáng qua.
Huyết áp đo tại khám: 145/90 mmHg; Xét nghiệm lipid máu: LDL-C 3.6 mmol/L.
Tăng huyết áp nguyên phát chưa kiểm soát tối ưu; Rối loạn lipid máu; Tác dụng phụ phù ngoại vi nghi ngờ do Amlodipine.
Research Contributor · Research Lab
Automated vulnerability detection and remediation for Infrastructure as Code, combining AST parsing with fine-tuned code LLMs.
Backend & CI/CD Platform Engineer · Product Team
Autonomous agentic recruitment system across 10 domain modules, automating enterprise resume ingestion, multi-modal signal extraction and candidate scoring.
ML Engineer · Independent Project
End-to-end tabular ML pipeline with 145 engineered features, SMOTE re-sampling, and XGBoost classifier deployed with Streamlit and Caddy.
Short case studies on the engineering decisions behind the numbers — what worked, why, and what the metrics do and don't prove.
LoRA fine-tuning Qwen3-4B made multilingual medical summaries fluent (54.59 chrF++). It took a separate validation gate with bounded retries to make them reliably correct.
Rule-based scanners judge a resource in isolation. Parsing 1,500+ Terraform files with tree-sitter and feeding the model each resource's real connections cut false positives by 28% versus Checkov and tfsec.
Filter by area or search by stack. Open a card to read its README inline.
6 projects
Clinical dialogue summarization across 17 languages with fine-tuned Qwen3-4B LoRA and mT5. 54.59 validation chrF++, a 5-layer evaluation framework, and a validation gate with bounded retries that lifted ConstraintScore from 85.9% to 99.5%.
Terraform security scanner that grounds a fine-tuned code LLM in tree-sitter AST context. 91.8% detection precision, 28% fewer false positives than Checkov and tfsec, with auto-generated patches in CI.
Autonomous recruitment system across 10 domain modules. Hybrid retrieval on 768-d multilingual embeddings with HNSW in pgvector, LangChain tool-calling for candidate evaluation, and Azure telemetry for token, latency and cost tracking with ABAC masking.
EcomoveX is a full-stack platform that supports eco-friendly navigation, sustainable travel habits, and environmental awareness. It helps users explore greener routes, complete sustainability missions, earn rewards, and view detailed place information powered by multiple integrated APIs.
ISC2
Specialization in Computer Networks & Telecommunications with a focus on Information Security. Academic Merit Scholarship, Jan 2025 and Jan 2026.
Entry-level security certification covering the five ISC2 CC domains.
Open to internships, research collaborations and conversations about applied AI or infrastructure security.
