QwikBuild

Inspiration

While there are quite a few coding agents available now, they are all meant for English-speaking, technically sophisticated users who are able to guide the agent. We aimed a bit higher: A coding agent that is available over mobile messaging, takes voice input, in any language, from non-technical users. This has the potential to make the creative side of computing more accessible to masses across the world.

What it does

It is an AI app maker agent that you interact over Whatsapp chat. You can send text, voice notes, images, any files - any input that is relevant. It leverages LLMs like Anthropic Claude 4.5 and uses QwikBuild's platform to delivery a working and deployed app - ready for public use in 7 minutes.

The platform provides the coding agent services for authentication, database, file processing and access to AI models. We also provide a dashboard where the user can manage and turn his app into an actual business.

For the hackathon, we added a major feature where taking the user's original AppCreationRequest message as input, a Claude-powered testing agent access the working web app, and comes back with automatic evaluation of its feature completeness and visual design. With this our agent can itself figure out what functionality is not working in the app as the user had originally requested!!

How we built it

We have two layers of agent: Support agent that handles intent extraction in WhatsApp and Coding agent. We also build an entire platform to provide modules like auth, database, file storage, AI gateway, comms gateway, and cron jobs.

For the testing agent, we leverage Claude 4.5 model with programmatic control of web browser instances in the cloud to perform functional and visual tests. The web page DOM is fed to the LLM which reasons and decides to interact - filling in form, clicking on buttons etc. We also use it for getting feedback on the visual design of the app.

Challenges we ran into

LLMs are becoming multimodal (Claude can take file inputs and can also call web-search and web-fetch tools), but LLM gateways typically do not support all the latest capabilities immediately. We found it better to use Claude APIs directly.
Many serverless platforms have tight limits on duration of computations. When a visual testing agent wants to take 50 steps over 10 minutes, the serverless functions can often time out.
While visual browser testing is available for watching as it is happening live, we currently have not finished implementing saving these screen recordings for later analysis.

Accomplishments that we're proud of

Starting from a voice note and getting a working and ready-to-use app in 7 minutes. The text/voice can be in any language!
Our two-layer approach helps with intepreting requests from non-sophisticated users to detailed instruction that are most effective for the coding agent to receive. This is what enables magic for most of our users.
Not just auth and database, but also file handling and AI tasks in the apps work most of the time
Our goal is not just to use the Testing Agent for ourselves, but actually build a test-fix loop so that user needs minimum involvement.

What we learned

Claude is magic.. and surprisingly cheap when consuming webpage DOM as tokens
Doing things in cloud is much nicer but long-running computation is becoming critical.

What's next for QwikBuild

Currently, the app testing agent is limited to the superadmins (us). We want to use it to enable automatic discovery and fixing of bugs whenever an app is created or modified.
We want to do a good launch.
We want to go from "Build an app" framing to "Build an online business" framing soon. This will require us to provide many more modules like payments, third-party integrations and so on.