AI Engineer at FPT Software AI Center with 2 years' experience building and deploying AI-powered applications — Large Language Models, Retrieval-Augmented Generation, Agentic AI systems, and machine learning model development.
Most of my work is taking agentic systems to production: agent runtimes, tool-use orchestration, permission and governance layers, and multi-tenant administration with auditing, usage and cost reporting. I design end-to-end ML solutions that bridge model training, backend engineering, cloud deployment and production support — model fine-tuning, optimisation, and latency-aware agent design.
I operate services through a modern observability stack (structured logging, distributed tracing, metrics, error triage, model/agent trace inspection) and work to disciplined engineering practice — enforced architectural boundaries, decision records, a default-deny permission model, forward-only database migrations, staged releases, and layered testing from unit through end-to-end and agent-behaviour evaluations. I regularly author technical specifications and review peers' work.
- 🎓 B.Sc. Computer Science — Applied Artificial Intelligence, HCMUT (VNU-HCM) · graduated Excellent, GPA 3.8 / 4.0
- 🌏 Vietnamese (native) · English — Professional Working Proficiency (IELTS 6.5)
- 🧭 Full-stack across FastAPI backends and React / TypeScript portals, deployed on AWS, Azure and Google Cloud
- 🌐 Full detail, live demos and every project → pdz1804.github.io
- Ho Chi Minh City University of Technology (HCMUT) — B.Sc. Computer Science, Major in Applied Artificial Intelligence ·
Sep 2022 – 2026Graduated with Excellent classification · Cumulative GPA 3.8 / 4.0 (official transcript, identifiers redacted) · Academic Incentive Scholarship (4/8 semesters) · OISP Scholarship (3/8 semesters) · Consolidation Prize, Bach Khoa Innovation Contest (Jun 2023) - Le Hong Phong High School for the Gifted — Mathematics Honours Class ·
2019 – 2022
- Top 10 outstanding projects, AI Riser Vietnam 2026 (Google for Developers · #BuildwithGoogleAI) ·
Oct 2026For Sách Của Em — a teacher studio that turns photos of textbook pages into printable, curriculum-cited lessons for Vietnamese primary-school teachers. Presented live at the Demo Day (GEM Center, HCMC) · ranked 4th of 10 in the audience-favourite poll (26% of the vote) · photos & certificate. - AI1 Young Talent, Level 3 (FPT Software) ·
Sep 2026 - "Best Team" award, FPT Americas (ST25) ·
2025
- Agentic ERP Platform (team of 15, stealth mode, ~1,000 users / ~100 DAU) — an enterprise AI workspace where autonomous agents act inside a company's business systems. Scope: agent runtime, permissions & governance, multi-tenant administration, integrations and customer-facing interfaces.
- Built the runtime permission layer (default-deny, enforced at runtime), per-customer administrator controls with a full audit trail, and usage & cost reporting; authored the permission-model specification through many review rounds.
- Shipped the first release of user-authored Skills and team sharing across agents, skills and artifacts; consolidated user & group management onto one source of truth.
- Built a third-party integration from scratch; implemented queued messaging; improved streaming chat (conversation search, composer usable while replies stream).
- Operate services via OpenTelemetry, Prometheus, Sentry and Arize Phoenix; daily CI work with quality and architecture gates and migration-safety checks; review peers on architecture boundaries, permission correctness and migration safety.
- Earlier: built the core agent harness of a healthcare agentic chatbot on AWS Bedrock AgentCore (team of 10) — >90% accuracy on a golden dataset built from customer discussions, typical response latency cut 30s → 15s → 8s — and solo-built its Management Portal. 🏆 Team received the "Best Team" award — FPT Americas (ST25).
- Presented three internal AI4ALL knowledge-sharing sessions — "AWS Strands Agents in the Cloud Era", "Portal for Agent Projects", "Agent as a Judge: How AI Evaluates AI" — and mentored an intern alongside senior engineers.
Remote supervisor: Prof. Kazuyuki Motohashi, The University of Tokyo
- Built a Blog System with AI search & recommendation on Azure AI Search, Cosmos DB (NoSQL), Redis and Azure OpenAI, using hybrid retrieval — BM25 + vector + semantic + freshness scoring.
- Designed M3ARAG, a locally deployable, GPU-accelerated Multi-Agent RAG system answering questions over PDFs, HTML, Office documents and text.
- Developed a Dual Attention Model extracting technical keywords from company websites, and a Transformation Matrix aligning Company–Patent data for the end-to-end Innovation Discovery pipeline.
- Designed an unsupervised framework to construct a Knowledge Graph with minimal domain-expert input, reducing manual labelling and curation effort.
- Evaluated SOTA LLMs for entity and intent extraction on raw Vietnamese documents; researched embedding, dimensionality-reduction and clustering techniques to reduce bias in abstract entity representations.
| Project | When | Stack |
|---|---|---|
| Agentic ERP Platform — enterprise AI workspace; agent runtime, permission layer, multi-tenant admin, Skills / Artifacts / Automations, plugin system | Jun 2026 – |
FastAPI · Temporal · PostgreSQL · GKE · React |
| Healthcare Agentic Chatbot & Management Portal — agentic chatbot at >90% accuracy (golden dataset) + knowledge / prompt / guardrail management portal | Nov 2025 – May 2026 |
AWS Bedrock AgentCore · Strands |
| Smart Product Recommendation System (SPR) — rule-based + LLM hybrid recommendation engine for a skin-health device; compliance, quality gates, automated testing | Oct – Nov 2025 |
Amazon Bedrock · Testing Automation |
| Sách Của Em — teacher studio: textbook photos → printable, curriculum-cited comics and explainers (AI Riser Vietnam 2026 Top 10) | Aug 2026 – |
Gemini · Firebase · Cloud Run |
| M3ARAG — GPU-accelerated multi-agent RAG for local, cloud-free document intelligence | Aug – Oct 2025 |
LangGraph · Docling · ColPali |
| Azure Blog System with Search & Recommendation — full-stack blog with hybrid semantic search and personalised recommendation | Aug – Oct 2025 |
Azure AI Search · Cosmos DB · Redis |
| Dual Attention Model for Innovation Discovery — attention model for technical-keyword extraction and Company ↔ Patent alignment | Jun – Jul 2025 |
PyTorch · Sentence-Transformers |
Academic · Fine-tuning LMs for NLP Tasks (T5 / BART / Flan-T5, LoRA) · Sentiment Analysis with Various Models (7 ML/DL models) · Detect AI-generated Text (DistilBERT vs. classical)
Personal projects, each with its own README. Descriptions are taken from the repositories.
| Repository | What it is | Stack |
|---|---|---|
| Tech News Mystery | Full-stack AI tech-news workspace: Crawl4AI / NewsAPI / Tavily ingestion, Qdrant semantic search, topic clusters on a PCA embedding map, a streamed chat agent and a voice agent (ElevenLabs STT/TTS over LiveKit with talk-over interruption); LangGraph on AWS Bedrock AgentCore, Celery workers, Terraform on AWS with GitHub Actions CI/CD | FastAPI · Next.js · LangGraph · Qdrant · Terraform |
| StreamSight | Real-time object detection and multi-object tracking on a 4 GB laptop GPU, with a WebSocket streaming console, a quantisation pipeline and an accuracy-throughput frontier | YOLO11n · ByteTrack · Python |
| EvalForge | Eval-first agentic RAG over the ML-arXiv literature: hybrid retrieval, grounded source-cited streaming answers, RAGAS quality gates that block regressions in CI | pgvector · BM25 · Redis · Next.js |
| AgentForge · FloraLens | A YAML-manifest agent-building platform (LangGraph ReAct runtime, tools / memory / guardrails, eval harness with regression gate) and a visual flower-similarity search built on its shared agent core | LangGraph · DINOv2 · Docker |
| Agent Harness & MRO Predictive Maintenance | An LLM-to-tool execution loop with human approval gates, and an aircraft-maintenance control desk, each with a narrated walkthrough video | TypeScript |
| Threadglass | A fast, local-first explorer for Facebook Messenger archive exports: browse, search and measure your own history without anything leaving your machine | TypeScript |
| AgentCore learning hub | AgentCore Runtime samples: a single Strands + Bedrock agent, and an MCP server + agent deployed as two runtimes | AWS Bedrock AgentCore |
| Agent-as-a-Judge deck | Bilingual (EN/VI) presentation and speaker materials from my internal AI4ALL talk, plus an agent-evaluation tooling reference | HTML |
| CS252 capstone | HCMUT capstone: multimodal lecture processing and a retrieval-augmented generation research platform | Python · AWS |
LLM & Agents — LangGraph · LangChain · AWS Strands · MCP (Model Context Protocol) · RAG & hybrid retrieval · Agent evaluation / LLM-as-a-Judge · OpenAI API · Google Gemini · Docling · ColPali
ML & Data — Hugging Face Transformers · Sentence-Transformers · PEFT / LoRA · spaCy · NLTK · Gensim · XGBoost · NumPy · Pandas · Polars
Platform & Ops — Temporal · Argo CD (GitOps) · Qdrant / ChromaDB / OpenSearch · OpenTelemetry · Prometheus · Sentry · Arize Phoenix · SonarQube · Backstage · Playwright
Full breakdown — proficiency & years
| Area | Detail |
|---|---|
| Languages | Python (advanced, 4y) · JavaScript / TypeScript (2y) · C / C++ (2y) · SQL (2y) · R (1y) |
| ML & DL | PyTorch · TensorFlow / Keras · Scikit-learn · XGBoost · Hugging Face Transformers (Datasets, Accelerate, Evaluate) · Sentence-Transformers · PEFT / LoRA & adapter fine-tuning · spaCy · NLTK · Gensim · Word2Vec · FastText |
| LLM & Agents | LangChain · LangGraph · AWS Strands · MCP · RAG systems · Agent evaluation / LLM-as-a-Judge · OpenAI API/SDK · Google Gemini · Docling · ColPali |
| Data & Analytics | NumPy · Pandas · Polars · Matplotlib · Seaborn · SciPy · Weights & Biases |
| Databases | PostgreSQL · MySQL · MongoDB · Azure Cosmos DB · Redis · Temporal · Vector DBs — Qdrant, ChromaDB, OpenSearch Serverless |
| Cloud & DevOps | AWS — Bedrock AgentCore / KnowledgeBase / Guardrails, EC2, S3, ECS, ECR, Lambda, CodeBuild, CloudWatch · Azure — AI Search, Cosmos DB, OpenAI, Blob Storage · Google Cloud — Cloud SQL, GKE, Artifact Registry, Cloud Storage · Docker · Kubernetes · Terraform · Argo CD · CI/CD · OAuth 2.0 / RBAC |
| Observability & Quality | OpenTelemetry · Prometheus · Sentry · Arize Phoenix · SonarQube · Backstage · Playwright / Selenium |
| Backend & Frontend | FastAPI · Pydantic · SQLAlchemy · Alembic · ReactJS · TypeScript · Tailwind CSS · Streamlit · Linux / Nginx |
| Practices | Architectural boundaries & decision records · default-deny permission modelling · forward-only migrations · staged releases · layered testing (unit → integration → e2e → agent-behaviour evals) · technical-specification writing · code review |
- Anthropic — Model Context Protocol: Advanced Topics, AI Fluency Framework & Foundations, Claude with Vertex AI (May 2026) · Introduction to Subagents, Claude 101 (Apr 2026)
- Google — AI Specialization, AI Fundamentals & 3 more (Feb 2026) · Prompting Essentials & 2 more (Jun 2026) · Foundations of Data Science (May 2026) · Gemini Certified University Student (Dec 2025)
- DeepLearning.AI — AI Agents in LangGraph · Functions, Tools & Agents with LangChain · LangChain for LLM App Development · Build AI Apps with MCP Server (+5)
- Hugging Face — AI Agents Fundamentals (Jun 2025) · DataCamp — AI Engineer for Data Scientists Associate (Sep 2025)
Google Cloud & the full list
- Google Cloud — Intermediate ML: TensorFlow on Google Cloud · Responsible AI for Developers · Inspect Rich Documents with Gemini Multimodality & Multimodal RAG · Develop Gen AI Apps with Gemini and Streamlit · Prompt Design in Vertex AI · Automate Data Capture at Scale with Document AI · Gemini for Data Scientists and Analysts
- DeepLearning.AI (full) — AI Agents in LangGraph · Functions, Tools and Agents with LangChain · LangChain for LLM Application Development · LangChain: Chat with Your Data · Build AI Apps with MCP Server · Knowledge Graphs for AI Agents · ChatGPT Prompt Engineering for Developers · Reasoning with o1 · Prompt Engineering with Llama 2 & 3
- AWS — Cloud Technology and Services Concepts · AWS Concepts
- Microsoft — Office Specialist: Excel, Word, PowerPoint
→ All credentials on LinkedIn.
Contribution graph, pinned repositories and language breakdown are on the profile below.
