Case study

Adalo

Overview

Elite Station contributes to the engineering behind Adalo’s AI-assisted app-building experience. Our work turns natural-language product requests into controlled, multi-step software delivery workflows: clarify intent, create an implementation plan, generate changes, validate them, and keep every action isolated to the right app workspace. The result is an AI coding-agent foundation designed for makers who need more than a one-shot prompt—they need repeatable, inspectable progress from idea to working application.

Technologies

AI agent platform: TypeScript, Node.js and Mastra-based coding agents for structured, tool-using workflows
Workspace model: Multi-tenant, organization-scoped Expo Router workspaces with per-app safety guards
State & observability: Redis-backed job and state management, durable per-app memory, and Langfuse tracing for quality, cost and latency analysis
Quality controls: Zod-validated tool actions, workspace guards, linting, type checks and automated tests

Our contribution

Agent orchestration: Architected and maintained multi-phase flows for clarification, specification, planning, code generation and validation—designed for per-app isolation in a multi-tenant maker platform.
Model routing: Built multi-model routing across Anthropic, OpenAI and open-weight models through OpenRouter, enabling deliberate model selection for different steps in the workflow instead of coupling the product to one provider.
Production controls: Designed typed, Zod-validated tool calls for file, schema and data work, backed by lint, type and test checks, workspace guards and parallel workers for concurrent screen generation.

Engineering outcomes

Cost and speed: The routing layer enabled a default-model change that reduced per-generation cost by approximately 30x and latency by approximately 3x.
Product learning loop: Durable per-app memory and Langfuse traces make it possible to inspect output quality, cost, token use and latency across model calls—creating the evidence needed to improve an LLM product over time.

Product capabilities

Natural-language requests are translated into a deliberate, multi-step implementation workflow rather than a single opaque generation step, giving makers more predictable results and teams clearer points for review.
Tool actions are validated and scoped to the correct workspace, helping keep multi-tenant generation safe, auditable and predictable.
Earlier work also included retrieval-augmented assistance over components, icons, fonts and help documentation, plus conversation persistence and support for 50+ languages.

Back to projects

Next project AlphaOnline

AlphaOnline is a longstanding automotive platform that offers services such as car valuation and online auctions for both new and used vehicles. Originally built in vanilla PHP, it has evolved into a robust custom framework developed in-house to meet the changing needs of the automotive market.

Go back to map