mAIntAIn

Inspiration

During my time living on campus and in student residences, I noticed that maintenance requests are a common part of daily operations. Property managers often receive numerous reports ranging from leaking pipes and faulty electrical fixtures to damaged furniture and broken facilities. These requests usually contain unstructured text, photos of varying quality, or even arrive through emails, making them time-consuming to review and prioritize.

I wanted to explore how AI could reduce this administrative workload by automating the repetitive parts of the maintenance process while still allowing managers to make the final decisions. This led to the idea of mAIntAIn, an AI-powered maintenance copilot that transforms maintenance requests into structured repair plans within minutes.

What it does

mAIntAIn automates the end-to-end property maintenance workflow. Residents can submit maintenance requests through a web portal or managers can import maintenance emails. The system then uses a multi-agent AI pipeline powered by Qwen Cloud to:

• Analyse uploaded maintenance images
• Validate image relevance and detect hazards
• Classify maintenance issues
• Assess repair urgency and priority
• Recommend suitable contractors
• Estimate repair cost and duration
• Generate communications for residents, managers and contractors

Before any action is taken, all AI recommendations pass through a human-in-the-loop approval stage, ensuring property managers remain in control.

How we built it

The project was built using a modular full-stack architecture.

Frontend:

• Next.js
• React
• TypeScript
• Tailwind CSS

Backend:

• FastAPI
• Python
• SQLAlchemy

Database:

• Supabase PostgreSQL

AI Layer:

• Qwen Cloud (Qwen3.7-Plus)

The backend orchestrates a sequence of specialised AI agents:

• Vision Agent
• Classification Agent
• Priority Agent
• Planner Agent
• Communication Agent

Each agent performs a single task and produces structured outputs that are passed to downstream agents, making the workflow modular and easy to extend.

The entire application was deployed on Alibaba Cloud ECS, using Nginx as a reverse proxy and systemd services to run the production frontend and backend.

Challenges we ran into

Building an end-to-end AI workflow involved much more than simply calling an LLM. Some of the biggest challenges included:

• Designing reliable multi-agent orchestration. We had to decide how information should flow between agents while keeping outputs structured and consistent.
• Ensuring reliable JSON outputs. LLMs occasionally returned malformed or unexpected responses, requiring robust parsing, validation and error handling.
• Handling multimodal inputs and edge cases. Real maintenance requests are messy—they may contain unrelated photos, mismatched text and images, missing images, or incomplete email reports lacking details such as unit numbers or clear issue descriptions. Building a robust system that could gracefully handle these scenarios required extensive validation, fallback logic, consistency checks, and specialised email processing before requests entered the AI workflow.
• Balancing automation with human oversight. Rather than fully automating contractor assignment, we designed a human-in-the-loop workflow so managers can review and approve AI recommendations.
• Deploying to production. Configuring FastAPI, Next.js, Nginx, systemd services, environment variables, reverse proxies and request timeouts on Alibaba Cloud ECS required significant debugging before achieving a stable deployment.

Accomplishments that we're proud of

I am proud that mAIntAIn is more than just an AI chatbot—it is a complete workflow automation system. Some highlights include:

• Successfully deploying the entire application on Alibaba Cloud ECS
• Building a modular multi-agent architecture powered by Qwen Cloud
• Supporting both web submissions and email-based maintenance requests
• Integrating vision understanding with structured maintenance planning
• Keeping humans involved at critical decision points through approval workflows
• Delivering a production-ready full-stack application within the hackathon timeframe

What we learned

This project taught us that building production AI systems involves much more than prompt engineering. I gained experience in:

• Designing modular AI agent architectures
• Orchestrating multiple AI services into a single workflow
• Engineering prompts for reliable structured outputs
• Building resilient APIs with proper validation and error handling
• Deploying full-stack AI applications on cloud infrastructure
• Designing AI systems that complement human decision-making rather than replacing it

What's next for mAIntAIn

Future improvements include:

• Mobile applications for residents and maintenance staff
• Integration with Outlook and Gmail APIs for automatic email ingestion
• Integration with enterprise property management systems
• Real-time contractor availability and scheduling
• Predictive maintenance using historical maintenance records
• Analytics dashboards for recurring maintenance trends and asset health
• Retrieval-Augmented Generation (RAG) over historical maintenance cases to improve repair recommendations and consistency across properties

Built With

• chatgpt
• fastapi
• git
• github
• next.js
• python
• qwen
• react
• sqlalchemy
• typescript
• vscode