Koder.aiTwo people shipped an AI app builder in six weeks, not quarters, and 70,000+ people in 100+ countries use it

1. Two agents that do a developer's daily work

   I designed the architecture around two agents. A planner coordinates the work, and an executor does the file operations: create, edit, search, lint, build. Every project runs in its own Docker workspace with a dev container and hot reload, and every version is kept in Git, so one user's broken build never touches another user's app and any mistake can be rolled back.

2. From an idea to a working app in 2 to 5 minutes

   The user describes the app, or starts from a screenshot or a Figma file. The agents write a React and Tailwind codebase, start the dev container and show a live preview within seconds, and a first working app is ready in 2 to 5 minutes. The same chat can go on to produce a Go backend with PostgreSQL and a Flutter mobile app. The user sees a working product before a traditional team would have finished the spec.

3. Code the customer owns, so lock-in never blocks a sale

   The platform has a code editor with a file tree, a visual editor, a database editor, snapshots with rollback and a planning mode. On paid plans the user can export the code or connect it to GitHub, which answers the lock-in objection that stops many buyers of app builders.

4. One click to publish, in the region the data rules require

   One click builds the app, uploads it to S3-based hosting and serves it through Caddy and a CDN, on a system domain or the user's own, with SSL issued automatically. I set up the infrastructure in several countries so an app can run where its data rules require, and on the Business plan the customer picks the hosting region. Customers meet data-residency rules without an operations team of their own.

5. Inference cost cut in half: several providers plus our own models

   I built the model layer, which routes requests across several model providers and our own models and inference, so no single vendor sets the platform's prices or its uptime. Our own models and the routing cut the cost of inference by 50%, which matters on a platform where every chat message is a model call. Enterprise customers can bring their own keys, use AWS Bedrock or Google Vertex AI, or run local models in their own data center.

6. Security ready for enterprise buyers from the first release

   I designed the security. All access to the app and the API goes over TLS, projects are separated by workspace and organization, access is set by role at the organization, workspace and project level, and key actions leave an audit trail. Enterprise plans add SSO/SAML and SCIM, and payments go through Stripe, so card details never reach the platform and there is no card data to lose.

7. Six weeks and two people to reach 70,000+ users in 100+ countries

   I also did the product design. A team of two built the platform in six weeks, a pace most in-house teams never reach. It now has more than 70,000 users in more than 100 countries, more than 35,000 apps have been created on it, and users add 500+ new ones every day.

Stack

• React
• Tailwind CSS
• Go
• PostgreSQL
• Flutter
• Docker
• Git
• Caddy
• S3 storage
• CDN
• Stripe
• AWS Bedrock
• Google Vertex AI