SPLEX 的 AI 應用程式建構流程為何不讓 AI 檢查自己的作品
Why our AI app builder never lets the AI check its own work
SPLEX 不讓編寫程式碼的 AI 檢查自己的作品,而由另一間 AI 公司的模型審閲每個檔案,再用真實瀏覽器操作應用程式並執行檢查。每次建構會提供安全性、測試、速度、完整度和易修改程度的評分;編寫前也會先詢問不清楚的需求。
When an AI writes code and then reviews that same code, it tends to approve its own mistakes. It "remembers" what it meant to write, so it reads what it intended instead of what is actually on the page.
We hit this constantly while building SPLEX (an AI app and website builder made in India), so we designed the whole pipeline around one rule:
The AI that writes your app never checks its own work.
Here's how that works in practice, and what it catches.
1. One AI writes, a different company's AI reviews
Every build goes through at least two models from different AI companies. One writes the code; another reads every file it produced. Because the reviewer was trained differently, it has different blind spots, and it flags things the writer glossed over.
A typical catch:
const age = parseInt(input);
The reviewer's note: "If someone types 'abc', this quietly becomes NaN and breaks the next step. Check it's a number first." The fix goes back to the writer before anyone sees "Ready".
2. A real browser clicks through the app
Reading code isn't enough. After the review, a real browser opens the app, clicks through it and runs its checks. Buttons that do nothing, pages that crash on load, forms that never submit: those show up here, not in your users' hands.
3. A quality score you can read
Every build gets scores for security, tests, speed, completeness and how easy it is to change, each out of 100, with the reasons. If something is weak, you know where to look.
4. Questions before code
Most bad builds start with a vague request. Before writing anything, SPLEX asks about whatever is genuinely unclear ("Who can see the bookings?", "Do prices include tax?") and never re-asks what you already said.
What's next: Fabo
For bigger, more complex software (SaaS products, booking and payment systems, internal tools), we're building Fabo, our coding model. It writes the tests first, runs everything in a secure sandbox, fixes what fails without breaking what already works, reviews security and logic, and hands you the code with a confidence score and a list of risks. It launches on 1 January 2027.
Try it
SPLEX is free to start at splex.in. Describe a website or app in plain words and watch it get built and checked. Paid plans start at ₹399/month, and new subscribers get 50% off the first month until 31 December 2026.
We'd love your feedback, especially on what the reviewer catches (or misses) in your builds. Drop a comment!
來源:DEV Community · dev.to