ML lead & AI architect
Hi, I'm
Viet Anh.
I build AI, and share what I learn.
I turn ideas into software people can use. This is where I share my projects, experiments, and lessons from making AI work in the real world — including the things that didn't go to plan.
I created AnyLabeling (3.5k★ on GitHub) and run Neural Research Lab.
Explore
Projects
AI tools, computer vision systems, and open-source work I have shipped.
View projectsAbout Me
My background, what I work on, and how I got into software and AI.
About meOpen Source
My work on GitHub: tools and libraries used by developers worldwide.
Visit GitHubBlog
Local-first AI, AI security, on-device Vietnamese voice, and what I am shipping next.
Read articlesLatest Articles
AI Can Write the Code. But Code Was Writing Us.
Writing software helped shape how we think as engineers. As AI takes over more implementation, I worry about discovery debt: the questions and insights we miss when we skip the work that might have revealed them. Here is how I want to preserve understanding, curiosity, and judgment while working with coding agents.
The GPU Was Already There: Four Silent Bugs in On-Device Training
Training is moving back onto the machines people own, and those machines are not a fleet you control. Apple GPU support sat on my plan for two years because nothing looked broken; the runs finished, they were just slower. Turning Metal on was worth 1.7x to 4.2x per epoch on the one M1 I own, and much less end to end. The measuring is what found the real bugs, three of which never raised anything: is_bf16_supported() answering True on a card that only emulates it, a gradient scaler silently dropping optimiser steps and halving a mAP, and 16-bit being slower than 32-bit on Apple Silicon even though everything works.
Ask, or It Will Guess: Question-Asking Is the Core Skill of Agentic Work
Models almost never ask you what you meant: across ambiguous questions, answer rates run above 95% and clarification barely registers. Repairing that in conversation makes it worse: underspecified multi-turn runs measure a 39% average performance drop, driven by a 112% jump in unreliability, and the effect shows up from two turns onward. The leverage is no longer in writing better prompts. It is in forcing the questions out before the first answer attempt, and in asking the questions afterward that prove you understood what you shipped.
I gave my website tools for AI agents with WebMCP, and thought hard about the attack surface
WebMCP lets a web page hand callable tools to browser AI agents instead of making them scrape the DOM. I wired three read-only tools into this site. The interesting part was not the API. It was deciding what an agent, or a prompt injection riding one, is allowed to do.
The Bottleneck Moved to Review: My SDLC After AI Writes Most of the Code
When a machine writes most of your diff, the constraint stops being how fast you type and becomes how well you review. Here is the SDLC I actually run (the provenance audit, the sandbox, the CI backstops, and the three questions I ask every AI-authored change) plus the parts of review that don't compress and never will.
Plan Once, Then Act: When the ReAct Loop Is the Wrong Harness for Small Local Models
On small local models, the standard ReAct loop has a failure mode nobody warns you about: the model calls one tool, declares victory, and stops. What we measured across 12 GGUF models in EdgeVox, why we added a plan-once dispatcher, and how to decide which loop your task actually needs.