AI builder, lifelong learner, product creator. 18-year advertising veteran learning to code at 40 and building AI products.
June: $57.01. July: $33.09. Two months after switching to DeepSeek V4 Pro via OpenCode, my AI costs dropped 75%. The commit history proves I shipped more, not less — 786 commits across 4 products. Here's the full breakdown.
I built an eval framework to grade my AI podcast pipeline across 40 real episodes in 7 languages. The review step was making scripts worse (3.34 → 2.91). Citation URLs all looked fake (score: 1.28). Chinese was the weakest locale (2.88). Six fixes, one day, and a new model later — here are the before/after numbers and every change I made.
I spent days crafting a system prompt for my AI podcast hosts. They read every word and ignored almost all of it. The fix wasn't better words — it was understanding that LLMs process rules differently depending on where they live in your pipeline.
The product I shipped in April was right about the engine. The product I settled on in June was right about the spine. In between: two complete curriculum rewrites, a framework that killed an architecture in a week, and the question I wish I had been asking since March.
I almost built a clever, AI-specific offline mode for DIALØGUE's iOS app. Then I asked a more boring question — what does iOS already give me? — and the real work turned out to be resisting the urge to be clever. Here's the standard, unglamorous machinery behind "download an episode and listen on a plane," and why boring was the senior choice.
I shipped DIALØGUE's iOS app as a port of the web product, then rebuilt it natively — three tabs, lock-screen audio, a synced transcript, resilient offline, and Siri — because a web app shrunk to a phone is still a web app.
TRANSMISSION served its purpose. Four and a half months later, time to move on. Over the Memorial Day long weekend I rebuilt the site, and the more useful story is the workflow: Claude Opus 4.7 did the design judgment, Codex on GPT-5.5 did the execution, and the /goal function let Codex run autonomously for close to four hours at a stretch.
Publicis has agreed to acquire LiveRamp for about $2.2B. I do not think the interesting question is whether this replaces the walled gardens or saves the open web. It does not. The better question is what advertisers still need outside closed ecosystems.
I spent 6 months trying to become a game developer while keeping my day job. I built three projects, wrote 684 commits, and shipped exactly zero games. Here is what happened, what I learned, and why nothing shipped.
I pushed what should have been an upgrade to my podcast platform's voice system. Six days and several commits later, I deleted 2,724 lines of code and rolled back to what worked. Here is what happened and what it taught me about testing production AI changes.
Nineteen days after cancelling Claude Max, my pattern has settled. Codex with GPT-5.4 on xHigh took the coding seat. Claude Code with Opus 4.7 on xHigh took every other seat at the desk.
People still message me asking if the 7 Andrew Ng courses I recommended in 2023 are the right path today. Short answer: most of them, yes, but with a different roadmap around them. Here is my 2026 update, with per-course verdicts and a forked path for builders and operators.
Prova started feeling real to me when the product began making real promises about progression, billing, auth, and state. The hard part was not generating code. It was building the contracts around it.
I keep seeing AI tools pitch agencies on content volume. But if you've ever managed real client relationships, you know the harder problem is trust: isolation, permissions, context, and not leaking one client's thinking into another's.
I thought I could splice together one course module, trim a few transitions, and call it a YouTube video. I was wrong. Building The Parade Problem taught me that good repurposing is not clipping. It is redesigning the idea for a different promise, a different audience, and a different first 30 seconds.
I cancelled Claude Max after 13 months and US$1,892.38 in subscription fees. This is not a victory lap. It is a 30-day test to see whether I can keep shipping STRATUM, DIALOGUE, my course platform, and this site at the same pace with Codex as my primary tool.
Two weeks after my original comparison, both tools shipped major updates. Codex challenged my product strategy in ways Claude Code did not. Claude Code shipped Agent Teams and AutoMemory. The result: I am cutting my $200/month Max plan — and getting better output for less money.
Most conversations about AI and team design start with headcount. I think that is the wrong starting point. The better question is which functions your team needs — and those turn out to be the same whether you have four people or forty.
DIALØGUE supports 7 languages, but the real multilingual work was not translating strings. It was fixing audience-local dates, TTS consistency, UI language drift, and deciding where quality mattered enough to slow down.
I spent years in advertising watching teams confuse motion with progress. Then I started building AI marketing tools and realized the problem was getting worse: faster execution, weaker judgment.
AI can now produce media plans, performance summaries, measurement frameworks, and campaign setups at impressive speed. The problem is not that the output is obviously bad. The problem is that it is often good enough to pass a casual review while missing the business context that actually matters.
Most teams still ask which model to use. From my experience, that is no longer the main question. If your AI system forgets the client or the brand, the category, and what good looks like, the smartest model in the world still starts every conversation from zero.
In 2023, I thought generative AI would flood search with cheap content and reduce the return on SEO. Three years later, that happened. But the bigger shift is that content production is no longer the moat. Structure, trust, QA, localization quality, and answer-engine visibility are.
One person. Seven modules. Three hours of video. Sixteen templates. A custom slide pipeline with 18 layout types. Professional voice clone. All while keeping my day job as VP. This is what the AI-first operating model looks like when you apply it to yourself.
The narrative that AI replaces entry-level work misunderstands what entry-level people actually do. In platform-heavy disciplines like activation, a junior isn't doing busywork — they're configuring targeting, QA-ing tracking, managing bid strategies. The real question is: when AI raises the floor for everyone, where does the advantage come from? Depth.
After using Claude Code with Opus 4.6 daily for almost a year, I spent a week with Codex and GPT-5.4. The verdict: neither tool wins outright. The combination — cross-model review, complementary strengths, operational resilience — is better than either alone.
Two weeks ago I wrote "Still building. Still not done." Today DIALØGUE is live on the App Store. Here's what the last 40% actually looked like.
493 blog posts across 17 years, translated into 10 languages, ~4,900 files, ~3.9 million words. Claude Code's parallel agents made it possible — but the Korean disaster, the Cantonese voice problem, and the 5-hour usage cap taught me more than the successes.
I'm building my first iOS app without knowing Swift. Claude Code scaffolded the whole thing in an evening. Then I opened the Simulator, and the real work began.
I migrated my blog to Next.js and thought the hard part was over. Then the compounding started — 6 mega guides, a smarter AI assistant, native newsletter, bot protection, and SEO overhaul in 8 days.
After visiting 26 national parks as an expat family, here's our honest guide — the 8 parks we'd revisit in a heartbeat, the ones that disappointed us, 4 road trip routes, and everything international visitors need to know in 2026.
From SSN delays to renting without credit, here's the relocation guide I wish I had before moving from Asia to the US.
From your first secured card to premium travel cards, here's the 4-stage system I'd use to earn rewards without getting sloppy.
How to buy T-bills, where to buy them, and when they beat a HYSA after state taxes if you live in California or another high-tax state.
From no file to 720+ in about a year, here's the roadmap I followed to build, protect, and use a US credit score as an expat.
The definitive 2026 guide to HSA, FSA, and HDHP for expats in the US—with updated IRS contribution limits, comparison tables, and practical tips from four years of navigating the system myself.
I migrated 485 WordPress posts to Next.js in 4 days using a plugin that taught Claude to think like a product manager—not just a code generator.
I built planning documents that taught Claude Code how to localize—not just translate—then watched it execute across 5 languages in parallel, turning literal phrases into native-sounding copy.
I rebuilt my podcast app in 14 days—a project that originally took 6 months. The acceleration terrifies me, and it should terrify you too.
I rebuilt my entire website in 3 days using AI—not as a coding assistant, but as my actual developer. I made the decisions; Claude Code wrote 28,000 lines.
I built 9 AI agents that forgot everything between conversations—wasting users 20-45 minutes weekly re-explaining their business. Here's how I made them share memory.
I shipped team collaboration features in alpha with <50 users because my customers were already working in teams—just through painful screenshots and copy-paste workarounds that cost 80+ hours per year.
I discovered my Postgres database had 89 foreign keys but zero indexes on them—turning millisecond queries into 843ms nightmares and nearly killing my alpha launch.
I found 31 blank screens in my SaaS—all because I forgot multi-tenancy isn't just about data access, it's about URL context. Here's how Claude Code helped me fix them all in one night.
I thought adding org_id to every table meant bulletproof multi-tenancy. Then my security audit revealed agencies were writing to SME tables—not through a bug, but by design.
After 20 years in agencies, I knew multi-tenant architecture couldn't wait—so on Day 2, with just one working AI agent, I tripled my dev complexity to avoid a future rewrite.
I fed my own landing page to my AI strategy agent and rebuilt it in 24 hours—conversion messaging transformed by dogfooding the product I'm selling.
I built a 9-agent marketing platform in 75 days that learns from every conversation—tell one agent about your business, and all nine get smarter together.
I spent 24 hours debugging why my React app kept making HTTP requests from HTTPS pages—even though my code was converting them. The culprit shocked me.
I built 3 AI marketing agents with multi-tenant architecture in 4 weeks—the same amount I accomplished in a month building my previous product that took 7 months to launch.