All posts

What I learned shipping 5 AI apps as a solo developer

I'm Marc de Ruijter, a Dutch full-stack developer with 25+ years in IT. I live in Chiang Mai, Thailand, and work remotely for clients in Europe. Since 2022 I've shipped 20+ products alone as Dizid Web Development.

This post covers five of them. Four call AI models; the fifth, TestSwap, makes no AI calls but was built with an AI coding assistant and for AI search. Here's what each taught me, failures included.

The five apps

1. Foreground: measure honestly

Foreground checks whether ChatGPT, Gemini, Claude and Grok name a business when its customers ask for a recommendation, and shows who gets named instead. Stack: Vue 3, TypeScript, Netlify Functions, Postgres on Neon, Stripe, and four AI APIs.

Foreground used to give a score from 0 to 100. In production it was useless: small businesses all landed between 0 and 12, while big names like Apple scored 57 to 100. No owner could move it. I replaced it with a plain count, "named in X of 12 answers", for questions from the customer's own niche.

AI answers change from run to run, so a change only counts at 3 or more of the 12, and "you improved" needs three scans in a row. One AI pass writes the action plan; a second checks it against the business's own website. In a test with 5 false claims planted in each of 5 plans, the check caught all 25.

2. TestSwap: AI crawlers don't run JavaScript

TestSwap lets indie developers test each other's apps: review one to earn a credit, spend it on a review of yours. Stack: Vue 3, TypeScript, Firebase Auth with X sign-in, Postgres on Neon, Netlify Functions. First commit: 9 October 2026. First real sign-in on the live site: the next day.

Crawlers like GPTBot and ClaudeBot don't run JavaScript, so a plain single-page app looks empty to them. The build prerenders the homepage, blog and legal pages into real HTML and writes an llms.txt. Every credit move is one database transaction, with rows locked in a fixed order so simultaneous reviews can't deadlock.

3. WuXing Zodiac: let payment be the gate

WuXing Zodiac is a Chinese astrology site covering all 60 element-and-animal signs, prerendered for SEO. Stack: Vue 3, TypeScript, vite-ssg, Netlify Functions and Neon. It sells a personal reading: Stripe Checkout, then a background function writes it with Claude, renders a PDF and emails it. The PDF is never stored.

That function's URL is public, so it trusts nothing but a paid Stripe session. It re-checks the payment with Stripe, claims the order in one database update and stops after 3 attempts, which caps the Claude calls per sale at 6.

4. LaunchPilot: count before you call

LaunchPilot is a launch checklist for indie developers: 33 tasks in 4 phases, with AI-written first drafts for each task. Stack: Vue 3, Tailwind CSS, Supabase, Netlify Functions, the Claude API and Stripe.

The browser never talks to Anthropic. A serverless proxy verifies the login token, takes the user ID from that token rather than the request, checks the plan's monthly quota, and only then calls Claude. A later review found no limit on input size; it now rejects anything over 100,000 characters.

5. Unplugged.cv: free AI endpoints need limits

Unplugged.cv is an AI CV builder I built in early 2026: paste your career history and get a CV and cover letter, with one rule for the AI: never invent employers or results. Stack: Next.js, React, Neon, Drizzle, the Claude API and Stripe. It is offline at the moment.

Its free endpoints called Claude without a login, as planned for the free tier, but also without a rate limit or a maximum input size. An automated scanner rated that fine; a code review caught it. The fix: 8 requests a minute per IP and an 8,000-character cap. Its downtime taught the other half: check the live site after every deploy, as TestSwap's smoke test now does.

What applies to every app

Cap AI spend in layers

A public form that calls an AI model is an open tap. I put a monthly limit at each AI provider, then limits in my own code: daily caps, per-IP limits, quotas per plan, input caps. Foreground also logs the cost of every call, and when a provider hits its cap, that AI drops out of scans and I get an email.

Stripe webhooks are where billing bugs hide

Foreground's webhook only updates the account whose subscription ID matches the event, so a late event for an old, cancelled subscription can't downgrade someone who subscribed again.

Keep operations boring

All five run on Netlify with serverless functions, four with Postgres on Neon. No servers to patch; that's how one person runs five products.

Use AI coding assistants, then check their work

I build with Claude Code. Each repo has a CLAUDE.md with rules the assistant must follow, such as never running a schema push on a shared database. It writes code fast; I still type-check, build, test and review anything that spends money. Those instruction files also go stale, so I review them like code.

Traffic is the hard part

In my first measurement, TestSwap was named in 0 of 12 AI answers in both its niches, and AI mixes it up with unrelated products of the same name. For WuXing Zodiac I first planned merch sold through paid ads; on paper that lost about $400 per 1,000 visitors.

Need an app built, or AI added to yours?

I build web apps and add AI features to existing ones: Vue or React, Node.js, Postgres, Stripe, and the OpenAI, Anthropic, Gemini and xAI APIs, with cost caps from day one. I work remotely for clients in the EU, at fixed prices, in Dutch or English.

See what I offer and what it costs, or tell me about your project.