AI Agents
Agentic workflows for real engineering work: Claude Code, Codex, spec-driven development and multi-agent orchestration.
Claude Code · Codex · LangGraph · Prompt engineering · Spec-driven dev · Multi-agent fleets
Luong Nguyen
@luongnv89
Agentic AI & Security Engineer
Senior Software Engineer at Montimage, Paris
I lead security and AI research projects by day and build open-source tooling for AI agents — including claude-howto, a 41k-star guide to Claude Code. Learn, build, share, connect.
01/About
Engineer first, driven by curiosity.
I'm a senior software engineer at Montimage in Paris, France, where I lead development on European research projects in cybersecurity, digital twins and applied AI — from deep packet inspection in C to fine-tuned security models and MCP servers. Fifteen-plus years in, I'm still an engineer who likes hard problems.
My edge is the overlap of AI and security: network-security engineering paired with AI developer tooling people actually use. claude-howto, my visual guide to Claude Code, has passed 41,000 GitHub stars, and asm, Milo and TextWiz are in developers' hands every day.
Engineering is only half of it. I work in a loop of learning, building, sharing and connecting — that's what keeps me going.
- Location
- Paris, France
- Role
- Senior Software Engineer, Montimage
- Experience
- 15+ years
- Focus
- Agentic AI · Agent skills · Security
- Research
- 7 publications · 9 EU projects
learn → build → share → connect
02/Focus
What I work on
Three threads, one loop: agents that do real engineering work, the skills that teach them how a team works, and the security research that keeps it all honest.
Agent Skills
Skills are how agents learn a team's way of working. I build the tooling to author, audit and distribute them across agents.
asm · Skill authoring · Skill evals · Security scanning · Claude Code · Codex · Cursor
Security
Security R&D at Montimage since 2014: deep packet inspection, LLMs for security, MCP servers and cybersecurity digital twins.
DPI · MMT · LLMs for security · MCP servers · Digital twins · Phishing defense
Method
Intention-Driven Development, run by agents.
Every piece of work starts as an intention, captured as a structured, agent-ready GitHub issue. IDD turns it into a work order; my skill set triages, resolves, reviews and self-checks it, and ships a tested PR. Chain that together and it runs for days without me in the loop — what friends call my software factory. Most of the apps below were built this way.
issue → analysis → resolve → review → self-check → PR
ToolboxPython · C/C++ · TypeScript · Swift · LangGraph · MCP · Docker · Kubernetes · CI/CD · Linux
03/Research
Security R&D at Montimage
Senior Software Engineer since 2014, leading development on European research projects in cybersecurity, digital twins and applied AI.
EU projects
COGNIMAN · SECASSURED · AIAgent4Cyber · NERO · CyberSuite · INTACT · GEIGER · PUZZLE · ENACT
H2020 · Horizon Europe · Digital Europe
LLMs for security
A fine-tuning pipeline for small security models: synthetic data, LLM-as-judge filtering, Unsloth/LoRA, and local deployment with Ollama.
MCP servers
Open-source servers on PyPI and the official MCP Registry that give AI assistants MITRE ATT&CK knowledge and threat-intelligence checks.
Deep packet inspection
Lead developer of MMT-DPI, Montimage's C packet-inspection engine (~80% of commits since 2015), including its 2026 hardening overhaul.
Phishing defense
A gamified awareness-training platform in 5 languages, an LLM-based email analysis service, and the GEIGER cyber range (900+ users).
Digital twins
COGNIMAN industrial pilot software, the secSIM cybersecurity twin platform, and the TaS IoT test and simulation platform.
Community
Organizing Committee for IEEE CloudCom 2026 and IFIP ICTSS 2026 in Paris; built both conferences' registration and payment platforms.
Selected publications
- ICTSS 2026
SecSIM: A Digital Twin Management System for Simulation-Driven Cybersecurity Assessment
- ARES 2025
- IFIP IoT 2024
SINDIT: A Framework for Knowledge Graph-Based Digital Twins in Smart Manufacturing
- QUATIC 2022
Digital Twin for IoT Environments: A Testing and Simulation Tool
- Book chapter 2021
Validation, Verification and Root-Cause Analysis — DevOps for Trustworthy Smart IoT Systems
- IEEE TIFS 2019
Reliable Detection of Interest Flooding Attack in Real Deployment of Named Data Networking
- IEEE WF-IoT 2014
DPWSim: A Simulation Toolkit for IoT Applications Using Devices Profile for Web Services
04/Products
Apps I've shipped
Designed, built and shipped end to end — click through to try them.
Milo
iOS · Voice AIHands-free AI for Siri & CarPlay — "Hey Siri, ask MILO…" routes your voice to GPT, Claude, Gemini, or 200+ models.
askmilo.pro
TextWiz
macOS · WritingLocal-first macOS app that runs AI "wizards" on selected text via a global hotkey (⌘⇧Space). On-device by default, no servers.
www.textwiz.pro
custats
macOS · iOS · AnalyticsClaude + Codex usage tracking in your menu bar — monitor token spend, sessions, and rate limits before they bite.
custats.info
devstats
Web · AnalyticsGitHub stats dashboard with contribution graphs, language breakdowns, leaderboards, and shareable profile cards.
devstats.info
music
Web · CLIA terminal-native music player with radio streaming and playlists — plus its web player and docs.
music-cli.luongnv.com
05/Open Source
Built in the open
Tools I build in the open — for AI agent workflows, developer productivity, and everything in between.
skillsSupercharge your AI agents/bots with reusable skillsPython★ 132
cc-context-statsKnow your zone. Ship with confidence — real-time Claude Code context usage monitoring across three operational zonesPython★ 117music-cliA command-line music player for coders — background daemon with radio streaming, local MP3s, and AI-generated musicPython★ 86
cclHit your limit? Need privacy? Just swap the model, everything else staysPython★ 39
sleek-uiModern, minimal UI component library — accessible, themeable, and designed for developers who want beautiful defaults without the bloatTypeScript★ 134
iddTurn GitHub Issues into structured, agent-ready work ordersShell★ 9
voice-castYour words, any voice. Voice cloning and text-to-speech with multiple TTS engines — clone any voice from a short audio samplePython★ 23
vscode-markdown-previewClean, minimal markdown preview for VS Code inspired by Zed — syntax highlighting, Mermaid diagrams, KaTeX math, and interactive featuresTypeScript★ 23Show 10 more projects+
repo-insightsGenerate comprehensive GitHub repo insights — real users vs bots, traffic, referrers, package statsShell★ 5
doc-basesIntelligent answers from any documentPython★ 4
bluetooth-diagnosticmacOS Bluetooth diagnostic tool — connection quality, signal strength, congestion analysis, and health scoring for all paired devicesShell★ 2
vscode-theme-neon-greenA vibrant neon green VS Code theme for developers who love the terminal aestheticJavaScript
video-to-gifConvert video (.mp4, .mov) to gif — simple, fast, and lightweightJavaScript★ 2
mdoctorKeep your Mac healthy — comprehensive CLI to diagnose, clean, fix, and benchmark macOSShell★ 6
git-auto-switchA CLI tool for managing multiple GitHub accounts with automatic identity switching based on workspace foldersShell★ 1
slide-speechAutomate presentations — play synced audio for each slide of a PDF or PPTX, timed to each audio file's durationPython★ 3
inbashBootstrap your dev environment in a single commandShell★ 1
llm-cliA wrapper of llama.cpp for easier use: find, download a model from Hugging Face, chat, get info, benchmark itShell★ 2
06/Writing
Latest writing
Notes on AI agents, security and the craft of shipping software.
- How I Use DeepSeek V4 Flash and GLM 5.3 Flash for Free (Every Day)
A step-by-step setup — free Cline account, 9router, and your favourite harness — that gives you 15M tokens per model per day of GLM 5.3 Flash, DeepSeek V4 Flash and Longcat 2.0, wired into Claude Code, Codex or Pi.
- How to Run Claude Code & Codex with Local Models via llama.cpp, Ollama, LM Studio, and vLLM (2026)
A 2026 hands-on guide to running Claude Code and Codex against local models served by llama.cpp, Ollama, LM Studio, and vLLM — with setup steps, trade-offs, and when to pick each runtime.
- How I Debugged Claude Code's 9% Cache Spike in One Prompt
One prompt consumed 9% of my session budget. By combining real-time usage tracking and per-interaction cache metrics, I traced the root cause to a 350K token cache recreation after a 13-hour idle gap — and asked Anthropic a hard question about their cache TTL design.
- How to Use gstack to Build a MVP Step by Step
A practical walkthrough of using gstack to go from idea to working MVP — covering the full workflow from planning to shipping with AI-powered skills.
- gstack Is Not a Dev Tool. It's Garry Tan's Brain on AI
Why gstack is more than a developer tool — it encodes Garry Tan's startup philosophy into an AI-powered workflow.
- Compatibility Debt Is Killing Perfectly Good Devices
How one Vite compatibility tweak revived an iPad mini 2, and why intentional legacy support still matters for real users.
07/Contact
Let's build something secure and useful.
Consulting, collaboration on AI agent tooling or network security, or just a hello — my inbox is open.
Paris, France · CET