ReuniteAI

ReuniteAI System Architecture Diagram
ReuniteAI Homepage
ReuniteAI Dashboard
Viewing an item detail from dashboard
Dashboard for found reports
notifications page in dashboard on ReuniteAI

Inspiration

The idea for ReuniteAI came from a feeling every student knows: the panic of losing something important. A laptop with your final project, your wallet, your ID card. On a busy campus like ours at the Federal University of Technology, Akure (FUTA), getting those items back often feels like a matter of pure luck. The old lost-and-found board just doesn't cut it. We saw a chance to turn this outdated process into an intelligent, proactive, and secure system that genuinely helps students, so we decided to build it.

What it does

ReuniteAI is a complete, AI-powered platform designed to automate the lost-and-found process for university campuses. It's a system of intelligent agents that work together to reconnect students with their belongings effortlessly. Key features include:

Effortless Item Reporting: Users can quickly report lost or found items by uploading a photo, providing a description, and pinpointing the location on an integrated map.
Automated, Proactive Matching: When someone reports a "found" item, our Search Agent automatically searches the database of lost items. If it finds a high-confidence match, the Notify Agent then sends a notification to whoever has reported the item missing.
Intelligent Hybrid Search: The Search Agent utilizes a powerful two-step hybrid search powered by TiDB Serverless. It first uses a fast full-text search to narrow down candidates by keywords, then runs a high-speed vector search on that smaller subset to find the most visually similar items.
AI-Powered Secure Claims: To prevent fraud, our Reasoning Agent [Moonshot kimi-thinking-preview model] acts as an AI security analyst. When a claim is made, it sends all the evidence (original reports, user's verification answer) to an LLM to make an intelligent judgment on the claim's validity.
Secure Pickup & User Dashboard: Approved claims generate a unique, one-time pickup code. Users can manage all their reports, track their status, and view notifications in a comprehensive and user-friendly dashboard.

How we built it

We built ReuniteAI using a modern, robust tech stack designed for an agentic AI workflow:

Frontend: HTML, CSS, and Vanilla JavaScript. This created a lightweight, accessible, and fast user experience without the need for complex build steps..
Backend: We built a scalable API with the Node.js Express framework and Clerk for protection of all endpoints.
Database: TiDB Serverless was our choice for its powerful hybrid capabilities. We heavily use both its full-text search for initial filtering and its vector search for accurate visual matching.
AI - Image Embedding: We deployed a custom BentoML service running the openai/clip-vit-base-patch32 model to convert images into 512-dimensional vector embeddings, or "visual fingerprints."
AI - Reasoning & Verification: The Kimi Thinking Preview LLM API is the brain of our Reasoning Agent, used to analyze evidence and make decisions during the claim process.
Image Storage: Cloudinary for reliable and fast hosting of all user-uploaded images.
Authentication: Clerk provides secure and seamless user sign-up, sign-in, and session management.

Challenges we ran into

Building a multi-agent system came with a few interesting challenges:

Execute SQL fail: Error 1105 (HY000): Faced this error and thankfully, Hazel was able to provide a fix, still trying to pinpoint where the issue erupted from tho.
Optimizing Hybrid Search: Getting our hybrid search to work efficiently in TiDB Serverless was a huge puzzle. We found that fts_match_word() had some limitations. Our solution was to develop a robust two-step query pattern: first, pre-filter candidate IDs with full-text search, and then apply the more intensive vector search only to that relevant subset. This made our searches both fast and accurate.
Crafting the Reasoning Agent's "Brain": Prompt engineering for our LLM-powered Reasoning Agent was a critical challenge. We had to carefully write a prompt that instructed the AI to act as a strict but fair security analyst. We needed it to approve legitimate claims while flagging vague or suspicious ones to prevent fraud.
Ensuring Proactive Logic: One of the magic of ReuniteAI is its proactive notification system. Architecting the backend to instantly and reliably trigger the search and notify agent whenever a new item is found required careful event-driven design to make the experience feel seamless and intelligent.

Accomplishments that we're proud of

We're incredibly proud that we built a project that nails the hackathon's theme of creating a multi-step, agentic workflow with real-world impact.

A Complete Agentic Workflow: We successfully built a full, end-to-end system of five distinct agents (Ingest -> Search -> Notify -> Reason -> Claim) . This automated workflow goes far more than a simple demo and provides a real solution to a common problem.
Proactive User Experience: We're especially proud of the Notify Agent. The system doesn't just wait for users to search; it actively works for them in the background and brings the solution directly to their attention.
Sophisticated Use of AI for Security: Our Reasoning Agent is a unique feature that uses an LLM for a complex judgment task, not just summarization. This adds a crucial layer of trust and security that a simple keyword-matching system could never provide.
Harnessing TiDB's Hybrid Power: Our two-step hybrid search is a powerful and practical demonstration of how to leverage TiDB's full-text and vector search capabilities together to solve a complex problem efficiently.

What we learned

This project taught us some valuable lessons:

Chaining focused agents unlocks real utility beyond single-call RAG.
Practical Database Implementation: We gained deep, practical insights into applying hybrid search with TiDB Serverless. Understanding the database's specific syntax and performance was crucial to building our core search feature.
AI is About Building Trust: We learned that a great user experience in an AI product is about more than just functionality. It's about building trust. By designing the system to be proactive (with notifications) and secure (with the AI-powered claim process), we created a tool that users can rely on.

What's next for ReuniteAI

We believe ReuniteAI has the potential to become an essential tool for universities everywhere. Our future plans include:

Deeper Integrations: Partner with university security departments to streamline the item handoff process.
Multi-campus tenancy and admin tooling.
Clarifying Q&A in reasoning loop to reduce manual reviews.
Real-time notifications (SSE/WebSocket) and email fallback.
Mobile app/updating mobile view.