MaintFlow AI

Inspiration

Our team has deep experience in the PropTech (Property Technology) industry. [cite_start]We know that for small startups trying to build maintenance marketplaces (e.g., for Airbnb hosts [cite: 244][cite_start]), the single biggest bottleneck is manual quoting[cite: 999, 1001].

[cite_start]A host reports a problem, and a skilled expert must manually analyze the issue, price materials, and estimate labor[cite: 1040]. [cite_start]This process is slow, expensive, and unscalable[cite: 998, 999].

We were inspired by the "Best Small Startup Agents" category. What if we could build an AI agent that acts as this skilled expert? [cite_start]An agent that could look at a host's problem description (jobDescription_raw) and photos (jobPhotos_raw) [cite: 65, 317][cite_start], and instantly generate a detailed, accurate quoteMatrix (our project's "v5" goal)[cite: 25, 108]? This AI would be the ultimate "force multiplier," allowing a 2-person team to do the work of 50.

How We Built It

We built a full-stack, "launch-ready" AI application using the hackathon's partner stack.

• Core Backend (LiquidMetal AI Raindrop): Our entire backend API and asynchronous job processing runs on the Raindrop Platform. [cite_start]We used Raindrop SmartSQL (Postgres) to store our users and jobs tables [cite: 30, 60][cite_start]—a pivot from our original NoSQL plan[cite: 4, 8].

• AI Model & Infrastructure (Vultr): We deployed a multi-modal vision model on Vultr Kubernetes Engine (VKE), leveraging their powerful cloud GPUs. [cite_start]This model is what reads the jobDescription_raw and jobPhotos_raw[cite: 65, 317].

• AI Integration (Raindrop SmartInference): We connected our Vultr-hosted model to our backend via Raindrop SmartInference. This gave us a clean, scalable endpoint for our agent.

• Storage (Raindrop SmartBucket): All user-uploaded photos (jobPhotos_raw) are securely stored in a Raindrop SmartBucket.

• "Launch-Ready" Features (WorkOS & Stripe): To meet the "Launch Quality" judging criteria, we implemented secure, enterprise-ready authentication using WorkOS. [cite_start]We also integrated Stripe to handle the full payment flow, from createPaymentIntent to payout[cite: 393, 395].

• [cite_start]Frontend (Netlify): Our client-side application (built with Flutter Web, based on our original wiki [cite: 403]) is deployed globally via Netlify.

Challenges We Faced

1. [cite_start]The Stack Pivot: Our original architecture was designed for Firebase[cite: 4]. Migrating our entire data model and backend logic to the Raindrop/Vultr/WorkOS stack in 35 days was our biggest challenge.

2. [cite_start]Structured AI Output: Getting the AI to return a clean, structured JSON for the quoteMatrix cite: 108 instead of just a text paragraph was incredibly difficult. This required significant prompt engineering.

3. Asynchronous Workflow: A user expects an instant quote, but our Vultr model takes 10-15 seconds to run. [cite_start]We had to re-architect our frontend to handle an asynchronous flow (polling the job status from pending_processing to bidding [cite: 329, 336]) for a smooth UX.

What We Learned

We learned that the "serverless AI" stack is incredibly powerful. The combination of Raindrop's Smart Components for the backend and Vultr for heavy-duty inference is a "best-of-both-worlds" approach.

Most importantly, we used the Gemini CLI (AI coding assistant), as required, to generate our SmartSQL schemas and API boilerplate, which sped up our development time by at least 30%. This project proved that AI agents are practical "force multipliers" for small teams.

Built With

• flutter-web
• gemini-cli-(ai-coding-assistant)
• liquidmetal-ai-raindrop-platform
• netlify
• postgresql
• raindrop-smartbucket
• raindrop-smartinference
• raindrop-smartsql
• stripe
• vultr-(vultr-kubernetes-engine)
• workos