Systems and processes, not one-off scripts.
I'm an AI engineer in Vietnam who deploys agents into real engineering work — not demos. My strongest work lives where AI meets messy reality: automating a legacy-to-microservice migration across dev and prod, building the agentic engine of a commerce platform, and taking a multi-agent banking system from zero to a national Top 7 in 48 hours.
I don't just use AI to code — I run a multi-agent engineering team around the clock: architect, backend, frontend, tester and security as governed AI roles under an operating process I wrote myself. Blocking quality gates, author-never-checks-their-own-work, and every recurring failure becomes a failing test. Background spans computer vision and cloud-native MLOps; I contribute upstream to OpenCode (174k★).
Agents, inside real engineering.
Four roles, one thread: taking work that was manual, messy and high-risk, and making an agent do it under gates a human can audit.
AI Engineer — Core Agentic Layer
- Built the automation engine on Claude Agent SDK + MCP — store/landing auto-builder, marketing, customer-support, support-email and cart-recovery agents running a D2C business
- Engineered agent safety at the harness level: PreToolUse guards that validate/block outbound actions (SMS, campaigns) before they fire, per-agent workspace path-guards, async audit logging on every tool call
- Built frontend, backend services and the admin orchestrator that coordinates & observes the agents
- Contributed fixes upstream to OpenCode — open-source agent runtime, 174k★
AI Automation Engineer
- Built an AI-agent automation system for legacy → microservice migration & QA across dev/prod — 56 APIs processed, 52 with full coverage
- Architected a 2-agent Processor + Verifier orchestration with anti-false-pass verification (verify-by-call, 3 runs/env) — human-in-loop caught 2 prod bugs the automated verifier passed
- Automated test-case generation from source: trace servlet → branch-coverage matrix → Postman/Newman → live verification
- Engineered multi-env data-safety gates: runtime credential resolution, capture-restore, prod write-protection
- Built a supervision dashboard (self-initiated, approved by the program owner): real-time migrate / test / build status per API with failure alerts — the lead runs the whole program from one screen
Independent AI / MLOps Engineer
- Shipped RAG + multi-agent platforms on Kubernetes (GCP/AWS), packaged so a new deployment is config-driven rather than a rebuild
- Ran the full MLOps lifecycle: orchestration, automated retraining triggers, model lifecycle as repeatable pipelines
- Built the infra + observability spine — Terraform, ArgoCD GitOps, Prometheus/Grafana/ELK — so failures surface before users hit them
AI Engineer
- Built and optimized production computer-vision models (classification, segmentation, OCR), owning the pipeline from data through deploy
- Operated live AI services in production — monitoring, retraining and incident response — plus internal R&D and technical presentations
Ten systems, one process.
Not ten side projects. The same written process — architect, backend, frontend, tester, security as governed AI roles — pointed at ten different problems. Here are eight of them plus the writing — what each one is, then the screen itself. About a minute.
Ten systems, not ten repos.
Each of these exists because a process needed to run without me watching it.
SHB Digital Expert Guild
A digital bank branch run by a team of AI experts (credit, legal, product, ops), judged live on stage. Approval gates enforced at the tool layer, a source reference on every number, full trace on every step. ~35k LOC, 202 commits in ~40h.
Anywe
A messaging network where AI agents plug in as contacts. Go control plane: signed-webhook delivery with retries & circuit breakers, fixed block vocabulary (agents can't draw their own approve buttons), append-only hash-chained ledger, atomic in-ledger settlement.
Agent Training Gym
Fresh agents run task batches, deterministic checks plus a **calibrated scorer (94% agreement with human verdicts)** grade them, batched edits pass a strict accept/reject gate, and the result is certified on locked test sets. Trained 5 banking-CRM skills to the domain's 85% bar; the framework caught 10 of its own harness bugs before they could corrupt a score. Runs on a subscription runtime — a ten-task round costs essentially nothing beyond ~3.5 minutes.
Life OS
Personal AI operating system — project / finance / automation tracing OS built in ~11 hours across 15 sprints by orchestrating an AI dev team under a self-authored process. 27k LOC, 1,089 tests, 12 modules, 0 live bugs.
Claude Manager
Observability + config management over Claude Code sessions: live tracing on SSE, cost and token-burn analytics, and 6 config managers (agents, skills, MCP, plugins, status line, hooks).
DevCrew
AI-agent team orchestrator — Kanban + chat + real-time SSE. Registry pattern: a new specialist team is 1 registry entry + 1 MCP + 1 prompt, **zero new code**.
Groundwork
"Jira for AI agents" — API test lifecycle (Postman + MCP), human-AI dual path (human owns setup, AI writes test logic), scoped agent tokens, audit trail, PII scrubber (AES-GCM).
OutboundOS
Five-phase pipeline (research → personalize → QA → classify → draft) with phase-gated evals. Business logic lives in agent prompts; Python is a thin tool layer — hot-swap subagent ↔ SDK with no call-site change.
PlatformDTC
The agentic engine of a Direct-to-Consumer commerce platform — the production system behind my current role.
LexiOps · Crawl2Insight
LangGraph multi-agent AIOps for Kubernetes (per-tool RBAC, continuous retraining) + a Vietnamese legal RAG; plus an end-to-end job-data platform on GKE with Terraform, ArgoCD, Airflow.
48 hours. A stage. Seven teams left.
Vietnam AI Innovation Challenge 2026 — NIC · Meta · SHB. Walked in, built a deployed multi-agent banking system in 48 hours, defended it live in front of the judges.
Top 7 — National Finals
Out of the full national field. Judged live on stage, not by submission.
~35k LOC · 202 commits · ~40h
A digital bank branch run by AI experts in credit, legal, product and ops.
Gates at the tool layer
Approval enforced where the action happens, a source chip on every number, full trace on every step.
Same process, different room
The multi-agent process I use daily, run under a 48-hour clock in front of strangers.
What I reach for.
Agents
- Claude Agent SDK
- MCP servers · design + ops
- Multi-agent orchestration
- Registry patterns
- LangGraph · RAG
- LLM judges · evals
Backend
- Go · chi · pgx/sqlc
- Python · FastAPI · Celery
- Bun · Hono
- PostgreSQL
- SSE · realtime
- Webhooks · HMAC contracts
Frontend
- Next.js · React
- React Native
- TypeScript
- Tailwind · shadcn/ui
- TanStack Query
- Block renderers
Infra / MLOps
- Kubernetes · GKE · Helm
- GCP · AWS
- Docker · Terraform
- ArgoCD · GitOps
- Prometheus · Grafana · ELK
- CI/CD · observability
Notes on building with AI.
Validate AI Workflows 50× Faster
A no-build-first method for scaling solo-founder efficiency — separating the experimental phase from production so you test concepts fast before committing engineering resources.
Agent-Building Is Training
Tools and skills are weights, your questions are the dataset, and the scorer is the loss function — why building an agent is a training run, and why the human holds the key.
Loop Engineering
How an AI team builds an app end to end without a human blocking the loop — kickoff, sprints, three verify gates, and the genealogy where every rule traces back to a real failure.
The Agent Gym
A training loop for agent skills: fresh agents on task batches, deterministic checks plus a calibrated scorer, batched edits through a strict gate, certified on locked test sets.
Awards & certifications.
- Top 7 (National Finals) — Vietnam AI Innovation Challenge 2026
NIC · Meta · SHB · AI for Vietnam — 48h build, judged live on stage - First Prize — 3rd Young Science Conference (IT Field)
Industrial University of Ho Chi Minh City · 2021 - IUH Scholarship
2019 — 2022 - Generative AI with Large Language Models
Coursera · 2024 - Machine Learning with Apache Spark
Coursera · 2024
Let's build something.
Open to **AI Automation / Solution Architect** roles, and to building agent systems that ship real value for engineering & ops teams.