Insights
Notes from building AI products
Decisions, trade-offs and measured results from the products on this site. Most notes are written in Vietnamese.
-
Governing a Claude Code rollout: caps, an off switch and CI cost on one Cloudflare Worker
An engineering org rolling out AI coding agents asked for seven controls: token visibility per developer, daily caps, spend alerts, a central off switch, a routing guard, CI failure cost and per-developer KPIs. What the vendor console covers, what it does not, and the small control plane I built to close the gap, with the numbers it produced.
-
Your work channels are an evidence index, not a scoreboard
Every mainstream productivity framework is team-level and warns you off individual evaluation. Managers still have to do it. A PM approach to reading Git, Jira, Confluence and Slack as evidence about a person, plus the design rules that keep it from turning into surveillance.
-
Distilling a RAG Knowledge Base: When Better Answers Come From the Corpus, Not the Retriever
My personal RAG had retrieval near its ceiling but was still giving truncated, sometimes-hallucinated answers. So I stopped tuning the retriever and rewrote the corpus itself: raw transcripts distilled into atomic, dated, self-contained claims. Answer quality went up with statistical significance on two separate corpora, and the fix removed hallucination instead of adding it.
-
In-App Purchases on App Store Connect: The Gotchas Nobody Warns You About
I set up In-App Purchases for a real App Store app for the first time and burned about 20 hours on failures that had nothing to do with my code. Five concrete traps - an agreement that looks signed but isn't, a save button that silently discards your work, staggered sandbox propagation, a subscription you can't test because you already own the product, and an icon rejected for one invisible pixel property.
-
BookTS: What Happened When an iOS App Went Through 100% of an Agentic Development Life Cycle
I built a 9-phase Agentic Development Life Cycle (ADLC) as reusable execution scaffolding, then ran a real App Store product - BookTS, an on-device EPUB-to-audiobook app - through all of it. Two of the phases caught bugs I would have shipped blind.
-
Building ADLC in Two Weeks: Using an AI Agent to Build the Harness It Runs Under
I paused a real App Store build to spend two weeks building the pipeline that would run it instead - a ten-phase Agentic Development Life Cycle, built by the same AI coding agent that would later execute it. What got reused versus built, the planning audit I overrode to own a review engine instead of just calling one, and the morning that engine caught a bug in its own predecessor's verdict.
-
ADLC: An Agentic Development Lifecycle for Solo, AI-Assisted Builders
Building alone with an AI coding agent removes every natural checkpoint a team gives you for free - no PM to catch scope creep, no reviewer to catch a dropped requirement. ADLC is the lifecycle apparatus I built to put those checkpoints back: ten phases, a gate discipline, and an autonomy contract that says exactly when the agent stops and when it doesn't.
-
I Built an AI That Audits My AI Infrastructure. Its Best Find Was My Own Blind Spots.
A week spent hardening fukuro - a personal auditor for AI infrastructure - turned into a lesson in recursion. I watched the auditor turn on itself, and its most useful output was a list of the boring, quiet mistakes that both I and the AI helping me had missed.
-
One Generator, a Mesh of Judges: An Idea-to-Ship Pipeline Built From AI Agents
Five AI agents wired into one idea-to-ship lane - one generator, four judges, and a runtime they share. The architecture, an end-to-end walkthrough, and the honest list of what it can't do.
-
Training a vision model to count objects in real time — running entirely in the browser
A PM journey: from 'the off-the-shelf API counts wrong' → fine-tuning YOLO → real-time on-device inference with WebGPU. Product decisions for building computer vision, not a code tutorial.
-
I Spun Up ~30 AI Agents in Parallel to Kill My Own Product Idea
Before building 'Jira for AI agents,' I ran one Claude prompt that fanned out into 111 sub-agent runs - roughly twenty minutes and 3.6 million tokens of work - across five research angles searched in parallel, 29 sources pulled, 133 claims extracted, then three independent skeptics voted on every claim. Only 9 survived. The verdict killed the idea in minutes. A PM's field notes on using AI to disprove yourself, cheaply.
-
Harness Engineering: The Scaffolding Is the Product, Not the Model
I ran two AI coding agents - Claude Code and Antigravity - through 100 turns of pair-programming over a Redis Streams channel I built, while they wrote real code against a live Redis. The code wasn't the lesson. What kept them working together was the harness. Field notes on engineering the scaffolding around an LLM.
-
AI-Canon-Crawler: why I gave my RAG a separate workspace for vendor truth
Personal-RAG kept hallucinating vendor specs because Marc's notes and Anthropic's docs lived in the same index. The fix: a dedicated canon workspace, a crawler that only ingests authoritative sources, and a routing rule that prefers canon when the question is about pricing, model size, or API behavior.
-
Knowledge-Audit for RAG-KB: From 80% to 99% with Grok 4.3
A real production use case: build a cross-source contradiction detector on Personal-RAG. Verify ground truth with Eval-Framework, swap Haiku → Grok 4.3, real accuracy 80% → 99%.
-
Mac-Translator: a Force Click translator that respects the flow
Building a menubar translator triggered by Force Click anywhere on macOS. Why the input gesture matters more than the model, and the AX/permission rebuild trap that cost me two evenings.
-
LLM bake-off across 6 judges: methodology + 7 surprising insights
Tested Claude (Haiku/Sonnet/Opus) vs Grok 4.3 vs GPT-5.4-mini vs Gemini Flash. N=30 dev + N=99 holdout. Stratified eval. Pairwise Cohen's kappa. Surprising winner.
-
7 decisions an AI assistant almost got wrong on an Eval-Framework SMB project
Field notes through a PM lens: building Eval-Framework for an SMB use case (~15k audits/mo). The AI engineering pair proposed 7 decisions that could have burned budget, leaked compliance, and shipped the wrong roadmap. PM catches + reframes. Counter-prompt playbook below.
-
Eval-Framework: why PMs need to measure before picking an LLM
Building a personal eval framework: from a 20-case smoke test → 192 stratified cases → a bake-off across 6 models. When 'guessing' which LLM is best gets expensive and wrong.
-
Diagram-Engine: text-to-diagram designed for AI agents, not humans
Why Mermaid/PlantUML break down when an LLM is the author. JTBD, syntax design constraints, and 37 layout rules that made the output good enough for a public playground.
-
Why your AI memory lies to you — and how I built a 3-layer Knowledge-Audit to catch it
Claude Code, Cursor, ChatGPT all ship 'persistent memory'. Six months in, the AI confidently asserts facts that are no longer true. I built a 3-layer audit plus a 4-tier cron to verify every claim.
-
Mail-Assistant: how I cut inbox triage from 90 min to 8 min with an ADHD-first AI agent
Field note: building a personal email triage agent that surfaces only what actually needs action. 568 messages → 7 P0/P1 rows. The 5 product decisions that killed 87% of the noise.
-
An AI sleep coach over Telegram: PM journey + 5 design principles for ADHD
I've worn a Garmin for 2 years - sleep, HRV, RHR data all there but zero action. This post logs my build through a PM lens: JTBD, existing alternatives, MVP scope decisions, ADHD-first design principles, and a V2 roadmap for a personalized AI sleep coach over Telegram.
-
Personal RAG KB: A 4-Version Journey from Cloud ADB to Local Mac M2 Max
Why I moved from Oracle ADB cloud → pgvector replica → local Mac - and what I learned measuring latency, quality, and cost at each version.
-
Building a Mac Voice Assistant and the Painful Lesson About Local LLMs
I built an AI voice assistant for daily use to truly understand the local vs cloud LLM trade-off. The surprise lesson: for agentic products, $0.001/turn Cloud is cheaper than 'free' local - once you count true TCO. A decision framework for product teams and CTOs.
-
5 Innate Biases of Cloud LLMs & 5 Levers to Overcome Them
Why Claude / GPT / Gemini get it wrong so often when writing workflow docs, PRDs, BRDs - and why you can't 'fix' the weights. You can only build guardrails around them.
-
The 15-Minute Deep Dive Summary: The Ultimate Prompt for NotebookLM
Turn any document into a deep, vivid 15-minute Podcast Audio Overview entirely in Vietnamese with this prompt.
-
Second Brain & LLM OS Model - Andrej Karpathy
Interactive page visualizing Andrej Karpathy's philosophy for building an LLM Operating System and RAG architecture.
-
Prompt Engineering 101 — The Art of Talking to AI
A roundup of fundamental and advanced Prompt Engineering techniques for interacting effectively with LLMs.
-
Welcome to Vuihoc.AI
First post - Introducing Vuihoc.AI and why I started journaling my AI research journey.