vibeDeploy - Zero-Prompt AI App Factory on DigitalOcean

Zero-Prompt Kanban: live GO/NO-GO ranking, build progress, and deployed apps — all streaming in real time
One-click Zero-Prompt Start launches autonomous idea discovery — no prompts, no setup, just click and watch

Inspiration

Most AI code generators stop at code. They produce files that fail type checks, skip deployment, and leave users to figure out hosting alone. We asked: what if an AI system handled the entire journey — from discovering the idea to deploying a live app — with zero prompts?

What it does

vibeDeploy is a Zero-Prompt autonomous pipeline on DigitalOcean.

One click starts AI-powered discovery from YouTube
Transcript extraction pulls source material from candidate videos
Gemini AI extracts app ideas and enriches them with academic research (OpenAlex, arXiv)
Competitive analysis checks market saturation
Deterministic scoring ranks each idea as GO or NO-GO on a live Kanban board
The user picks any GO card to build
Contract-first code generation produces type-safe frontend + backend
Build validation iterates until quality passes (target ≥ 85% match rate)
Automated deployment ships the app to DigitalOcean App Platform with a live URL and public GitHub repo The entire flow — from discovery to deployment — runs autonomously. The user only makes one decision: which GO card to build.

How we built it

vibeDeploy is a dual-runtime application:

Backend: Python 3.12, Gradient ADK, FastAPI, LangGraph state graphs
Frontend: Next.js 16.1, Tailwind CSS, Framer Motion, SSE for real-time updates
Infrastructure: DigitalOcean App Platform, Managed PostgreSQL, Spaces The architecture is contract-first: an OpenAPI spec is generated before any code, then TypeScript types and Pydantic models are derived from it. Code is generated per-file with targeted retry on failure, validated through syntax, import, and build checks, then deployed with health verification.

DigitalOcean Gradient™ AI usage (13 capabilities)

Gradient ADK — agent entrypoint and orchestration
Knowledge Bases — contextual retrieval
Evaluations — output quality scoring
Guardrails — content safety
Tracing — pipeline observability
Multi-Agent Routing — council and pipeline coordination
A2A Handoff — agent-to-agent delegation
Serverless Inference — LLM calls via DO Inference API
App Platform — dual-component deployment (api + web)
Spaces — artifact storage
Image Generation — visual assets
Agent Versioning — deployment history
MCP Integration — tool connectivity Managed PostgreSQL stores sessions, cards, lineage, and workflow state.

Challenges we ran into

Session persistence: the orchestrator stored sessions in memory, losing all state on server restart. We implemented DB-first hydration with lazy loading from PostgreSQL.
LLM-generated lockfiles: AI-produced package-lock.json files caused npm ci mismatches on DigitalOcean's Node.js buildpack. Fixed by stripping fake lockfiles when real npm validation is unavailable.
Stream termination: LangGraph's astream_events didn't reliably signal completion after the deployer node finished. We solved this with mid-stream card status updates — detecting deployment events during streaming rather than waiting for the loop to end.
Build quality: single-shot code generation had ~40% deploy success. The contract-first, per-file approach with iterative validation raised this to ~90%+.

Accomplishments we're proud of

A fully working Zero-Prompt pipeline: click → discover → rank → build → deploy
Live deployed apps generated autonomously with working URLs and public GitHub repos
Real-time Kanban UI that streams every step as it happens
13 DigitalOcean Gradient™ capabilities integrated into one coherent workflow

What we learned

Validation matters more than generation quality — bad code that gets caught is better than good code that silently fails
Contract-first architecture (OpenAPI → types → code) eliminates most frontend/backend mismatches
Live visibility into AI workflows (streaming Kanban, action feed) builds trust and enables debugging
Deployment is not the last step — it's an integral part of the feedback loop