AI Alone Wasn't Enough – They Added Humans and Raised $6M

01 /The Concept

Building a mobile app without writing a line of code used to mean settling for something that looked like it was built without writing a line of code. Woz is trying to change that equation.

The platform generates production-quality mobile applications from plain-language descriptions – targeting small businesses, solo entrepreneurs, and creative professionals who either lack the technical skills to build themselves or the budget to commission developers externally.

- In the first step, the author describes the desired functionality in plain language – the platform generates the app's code.

- The author then tests the app, confirms it works as intended, and checks that nothing breaks during normal use.

- Once the author gives the green light, the platform does a final pass, adds finishing touches, and helps publish the app to the App Store.

- When the author wants to add features, change the interface, or make any other modification, they return to the platform – which produces a new version and helps publish it.

After the initial launch, the platform starts running "shadow" maintenance autonomously: monitoring hosting health, handling routine cleanup and optimization, catching and fixing bugs, managing security, collecting and surfacing analytics, and more.

Woz doesn't just generate the code – it takes on the entire process of deploying and operating a live app. Most non-technical authors don't anticipate how complex this is, and most AI coding platforms treat their job as done the moment the code is generated.

The internal development process at Woz is also structured distinctively. The startup calls it an "anti-vibe" approach – a deliberate contrast with other vibe coding platforms.

The idea: the platform first decomposes a task into individual subtasks with clear specifications, then assigns each subtask to a separate AI agent purpose-built for that type of work. The app is assembled like an industrial production line – where each station performs one operation, extremely well, almost automatically. The startup calls its platform a "factory" for making apps, not a vibe coding tool.

Woz was [covered here](/review/gde-vzjat-tehnicheskogo-kofaundera) at the end of last year, when it had just entered Y Combinator. It graduated in the spring, and has now raised a new $6M round.

02 /Why It Matters

The approach described above is already interesting on its own.

But the most interesting development is the updated platform's operating principle: "Build what you can, but bring in experts when AI hits its limits." In other words, Woz has moved to a hybrid model that combines AI with live expert developers.

Most people who turn to popular vibe coding platforms like Lovable or Replit expect them to produce flawless apps. Reality is considerably harsher: the quality of AI-generated code is often far from polished, leading to bugs and unpredictable behavior. And those problems in AI-generated code can typically only be caught and fixed by experienced engineers whose natural intelligence still exceeds the artificial kind.

A revealing data point: traffic to Lovable, Bolt.new, and Vercel v0 has dropped 30–60% from their peak in June. And prior studies on the quality of code generated by ChatGPT, GitHub Copilot, and Amazon CodeWhisperer found that they produced correct code in only 30–60% of test cases.

So the new, fundamentally important addition to Woz's platform is this: before the generated code reaches the author, live expert reviewers check it.

Those experts are doing what's called code review in professional software development – a step where code written by one developer gets examined by another before it goes live. Here, the experts review AI-generated code rather than human-written code, and their feedback also feeds back into improving the AI's algorithms.

The result: code that no longer resembles the "spaghetti code" produced by many vibe coding platforms. It looks like code written by professional developers – structured, clean, readable, and secure.

The founders plan to extend this same industrial, hybrid approach to building web services, smart TV apps, smart glasses apps, and other platforms. And importantly, to applications in heavily regulated sectors like healthcare and insurance – where the cost of a bug is too high to hand the keys entirely to AI, but where AI could dramatically accelerate and improve the development process.

The standard tier, in pure AI mode, costs $49 per month – though each code export costs an additional $100.

The tier with live expert code review costs $499 per month – currently available in limited beta.

03 /Opportunities

The unbounded euphoria about AI's limitless capabilities appears to be slowly deflating as real constraints come into focus – constraints that may never fully disappear, even as they evolve with improving AI.

At Y Combinator's Demo Day in the spring, Woz was boasting that its AI platform would learn to write anything well enough to fully replace a technical co-founder.

In practice, Woz evidently ran into fundamental limitations that led them to introduce human review and editing of AI-generated code. And this doesn't look like a temporary workaround – it looks like a durable "human-AI hybrid" concept.

The main trend and direction: applying similar hybrid approaches to professional-grade platforms built for real tasks across different domains.

A couple of examples from other sectors show the same pattern.

Counsel ([related review](/review/novaja-model-dlja-marketplejsov-uslug)) just raised $25M on a medical consultation platform where users interact with a specialized AI but can switch to a live physician whenever they want or need.

Crosby ([related review](/review/bolee-prostaja-model-dlja-sozdanija-perspektivnogo-ii-produkta)) raised $20M in early October on an "AI legal firm" for reviewing and fixing contracts that clients receive from their counterparties. AI does the heavy lifting; human attorneys verify results and handle the subtle edge cases the AI missed.

One interesting wrinkle: as these platforms scale, the startup's own expert pool may not be sufficient. They'll inevitably need to bring in external freelancers – which effectively turns them into marketplaces embedded inside an AI platform.

This points to another trend: the transformation of traditional service marketplaces from "directory of business cards" into AI-first platforms where AI delivers real value to users while simultaneously lifting the routine workload from the service providers.

So there are two distinct entry points into building these kinds of platforms:

- Look at popular "pure AI platforms" for specific tasks – and figure out how much quality improvement is possible by adding human specialists to refine inputs, oversee the process, and improve outputs.

- Look at what services people search for online and on service marketplaces – and figure out where and how to insert AI into the ordering and delivery process to raise quality and reduce the routine burden on the people providing those services.

So – from which direction, and into which service sector, would you want to enter with this model?