P.01 — AI · Full-Stack
Lumos AI Application Builder
A multi-agent application builder that turns natural-language requirements into architecture, UI/UX plans, code, integrations and deployable applications.
- Role
- Associate Software Engineer · Primary author
- Year
- 2025 — 26
- Company
- CodeGen International · Lumos
01 — Overview
Overview
Lumos is CodeGen's agentic AI R&D team — around ten people with a flat structure. The application builder is its most ambitious product: describe an application in plain language and a pipeline of specialised agents takes it from requirements to a deployable build.
It ships as an Electron/Python desktop client and a Next.js web app, and has been demonstrated to several leading software companies in Sri Lanka, with rollout to their engineering teams planned.
Started from zero. I wrote about 80% of the modules I own, with contributions from the Lumos team.
02 — The problem
Building applications from natural language requires far more than generating code.
Requirements have to become an architecture, the architecture has to become interfaces and data models, and the result has to integrate, validate and actually run. A single prompt-to-code step breaks down long before that.
03 — The system
From requirement to running application
Each stage is owned by a specialised agent with its own tools, structured outputs and validation — with retries and fallbacks when a step fails.
- 01
Requirements
Natural-language intent is clarified into structured requirements.
- 02
Planning
Scope is broken down into features, flows and delivery steps.
- 03
Architecture
Agents propose the application architecture and data model.
- 04
UI / UX
Screens and interaction plans are generated from the architecture.
- 05
Implementation
Code and database integrations are written stage by stage.
- 06
Validation
Structured output validation, retries and fallbacks keep it on track.
- 07
Deployment
The result is packaged as a deployable application.
04 — My contribution
Team, and what was mine.
Team
- Lumos R&D team of about 10, flat structure — everyone contributes
- Architects, team leads and VPs for direction and review
My contribution
- Started the platform from zero as primary author
- Built the multi-agent pipeline with LangGraph, LangChain and CrewAI
- Tool calling, structured output validation, retries and fallbacks
- Electron/Python desktop client and Next.js web app
- Technical demos to external engineering teams and clients
05 — Technical depth
What it's built with, layer by layer.
- AI
- LangGraph · LangChain · CrewAI · OpenAI · Anthropic Claude · Google Gemini
- Agents
- Multi-agent orchestration · Tool calling · Structured outputs · Prompt & context engineering
- Clients
- Electron · Python · Next.js · React
- Infrastructure
- Microsoft Azure
06 — Gallery
In pictures.
Internal platform — product screens are under NDA, so these images set the scene rather than show the UI.
07 — Outcome
Where it landed.
7
Pipeline stages
each owned by a specialised agent
2
Clients
desktop (Electron/Python) and web (Next.js)
Demonstrated to several leading Sri Lankan software companies, with rollout to their engineering teams planned.
Part of Lumos, which received a special recognition at the National AI Awards 2026 in the Best AI Solution category.