Technical Breakdown of Campus Find: A DeepSeek-Powered AI Lost & Found System

Three Major Pain Points of Traditional Lost & Found

Losing personal belongings is a common frustration in campus life. Yet traditional lost and found methods suffer from three significant pain points: fragmented information channels — people don't know where to post when they lose something, leading to severely scattered information; lack of intelligent matching — users can only manually scroll through listings one by one, making the process extremely inefficient; and cumbersome and insecure offline handoffs — there's no trustworthy communication bridge between the person who lost an item and the person who found it.

At the 8th Chuanzhi Cup (National Finals) AI WeChat Mini Program Development Challenge, a team presented a lost and found mini program called "Campus Find" (校友寻). By integrating DeepSeek's series of AI models, it fundamentally addresses these pain points. This article provides an in-depth analysis of the mini program's design philosophy and implementation from three perspectives: technical architecture, core features, and AI capabilities.

Technical Architecture: A Full-Stack Solution with Separated Frontend and Backend

Frontend Presentation Layer

"Campus Find" is built using the native WeChat Mini Program framework, with a tech stack comprising JS, WXML, and WXSS, along with integrated LBS geolocation services and OCR image recognition. The native WeChat Mini Program framework is a development framework purpose-built by Tencent for the WeChat ecosystem. WXML (WeiXin Markup Language) is an HTML-like markup language, WXSS (WeiXin Style Sheets) is a CSS-like styling language, and together with JavaScript they form a complete frontend development system. Compared to cross-platform frameworks like Taro or uni-app, the native framework delivers better performance and more comprehensive API support within the WeChat ecosystem — especially when calling native capabilities like WeChat login and geolocation. LBS (Location-Based Service) is one of the core capabilities of mobile internet. It obtains user geolocation through GPS, Wi-Fi, cell towers, and other methods, providing the technical foundation for location tagging and nearby search in lost and found scenarios.

The frontend also integrates the Amap (Gaode) Maps API for precise location selection and the QWeather API to provide users with real-time weather information.

For performance optimization, the project employs subpackage loading, which loads feature modules on demand and effectively reduces the initial screen loading time. WeChat Mini Programs have strict size limits on the main package (no single subpackage exceeding 2MB, with a total size cap of 20MB). Subpackage Loading allows developers to split a mini program into multiple packages — only the main package is downloaded when the user first opens the app, while other feature modules are downloaded on demand when actually accessed. This mechanism directly impacts user experience — research shows that for every additional second of initial load time, user drop-off rates increase by approximately 10%. For a feature-rich application like "Campus Find," splitting non-essential first-screen features such as the forum, AI mascot, and rewards redemption into independent subpackages is a critical optimization strategy for ensuring fast startup.

Backend Service Layer and Deployment

The backend uses a Node.js + Express.js architecture, with MySQL as the relational database for the data persistence layer, following 3NF (Third Normal Form) design principles. Node.js is a JavaScript runtime built on Chrome's V8 engine, and its event-driven, non-blocking I/O model makes it particularly well-suited for handling high-concurrency network requests. Express.js is the most popular web application framework in the Node.js ecosystem, providing core features like route management, middleware mechanisms, and template engines for rapidly building RESTful APIs. 3NF is an important standard in relational database design that requires every non-key attribute to be neither partially dependent on nor transitively dependent on candidate keys. In simple terms, 3NF ensures data consistency by eliminating redundancy — in the lost and found scenario, user information, item information, and location information are properly separated into different tables and linked through foreign keys, avoiding update anomalies caused by storing the same data in multiple places.

The entire project adopts a separated frontend-backend architecture, with the backend independently deployed on a VPS server running Windows Server, accessible externally through port mapping.

On the AI side, the project integrates multiple DeepSeek models: DeepSeek R1 for deep reasoning and intelligent analysis, and DeepSeek V4 Pro combined with DeepSeek for the AI mascot interaction. A complete AI persona system is established through API-Key deployment and Prompt Engineering. Prompt Engineering is one of the core techniques in current large model application development — it refers to the methodology of carefully designing input prompts to guide AI models toward desired outputs. In the "Campus Find" AI mascot system, Prompt Engineering serves to establish a stable "persona" for the AI, including role definition (campus lost and found assistant), language style (friendly and approachable), capability boundaries (only answering questions related to lost and found), and safety constraints (not disclosing user privacy). A well-crafted System Prompt typically includes role definition, task description, output format requirements, few-shot examples, and safety guardrails. By calling DeepSeek's API through an API-Key and pairing it with carefully tuned prompts, developers can build AI applications with domain-specific capabilities without training a model from scratch.

Core Features in Detail

Smart Posting and Information Management

The posting page is thoughtfully designed, featuring modules for post type, basic item information, item description, location information, and contact details. The system enforces strict validation on required fields — the submit button appears grayed out when required fields are incomplete and turns yellow (clickable) once everything is filled in.

For location selection, the Amap API is integrated, supporting both automatic positioning and manual search. The time selection module also includes logic validation — an error prompt is triggered if the end time is earlier than the start time. The contact information section similarly validates phone number length to ensure accuracy.

Each published post also supports edit, delete, and boost operations. The boost feature pins a post to the top of the homepage for 30 minutes, helping the owner gain more visibility during the critical time window.

Account Security and Session Displacement Mechanism

The project features a unique "session displacement mechanism" for login security. Session displacement (also known as "login kick" or "single sign-on mutual exclusion") is a common account security strategy whose core principle is that the same account can only maintain an active login session on one device at a time. This mechanism is widely used in instant messaging and social applications (such as WeChat and QQ's PC login notifications). Its technical implementation typically relies on a Token mechanism: the server generates a unique Token bound to the device upon login, and when a new device logs in, a new Token is generated while the old one is invalidated.

When the same account logs in on another device, the original device receives a login alert, and the user can choose either "Not me" or "Confirm it's me." Selecting "Not me" immediately bans the account, while "Confirm it's me" allows the new device to log in and signs out the original device. The "Not me" one-click ban feature that "Campus Find" adds on top of the standard mechanism is a security enhancement tailored for the campus scenario — if an account is compromised and used to post fraudulent found-item listings, the original user can immediately freeze the account to prevent further fraud.

Additionally, the login page includes a custom human verification feature that uses encapsulated click events to filter out some automated script attacks.

Built-in Private Messaging and Social System

The mini program includes a complete private messaging system supporting text, images, and emoji. Messages are transmitted through Node.js, stored in the MySQL database with precise timestamps. The messaging page also distinguishes between friend and non-friend message lists, and supports social features like friend requests and message deletion.

QR Code Confirmation for Returns: A Closed-Loop Handoff Design

This is one of the most practical closed-loop designs in the entire system: each published post generates a corresponding QR code. When the owner and finder meet in person, they can scan the QR code through the mini program to confirm the item handoff. The handoff record is permanently saved, and the homepage counters for "Successfully Recovered" and "Helpful Returns" are automatically updated, creating a positive feedback loop.

AI Capabilities: The DeepSeek-Powered Intelligent Core

Deep Reasoning and Intelligent Matching

The project leverages DeepSeek R1's deep reasoning capabilities to intelligently analyze multi-dimensional information about lost items — including scenario, time, and location — automatically matching lost and found listings and precisely pushing potential matches. DeepSeek R1 is a reasoning-enhanced large language model developed by DeepSeek, and its core feature is a human-like "slow thinking" capability. Unlike traditional LLMs that generate answers directly, R1 performs multi-step Chain of Thought analysis before responding — progressively breaking down problems, verifying hypotheses, and drawing conclusions. This capability is especially critical in the intelligent matching scenario for lost and found — when the system needs to determine whether a "black backpack lost on the third floor of the library" matches a "dark gray backpack found near the teaching building," R1 can comprehensively analyze color similarity, geographic distance, time window overlap, and other multi-dimensional factors to produce a confidence score, rather than relying on simple keyword matching. This is the key technology for solving the traditional "matching difficulty" pain point.

OCR Image Recognition with Auto-Fill

Users can simply take a photo to automatically identify item information, extract item features and text, and generate a standardized lost item description with one click. OCR (Optical Character Recognition) is a classic technology in computer vision whose core task is converting text in images into machine-editable text data. Modern OCR has evolved from early template matching to deep learning-based end-to-end recognition, capable of handling challenging scenarios like complex backgrounds, multi-angle shots, and handwriting. In a lost and found application, OCR's value lies not only in recognizing text on items (such as names and student IDs on ID cards, or titles on book covers) but also in converting unstructured image information into structured data, enabling AI to perform further semantic understanding and intelligent matching. For example, when a user photographs a campus card, OCR can automatically extract key fields like name, student ID, and department, eliminating tedious manual input and significantly lowering the barrier to posting.

AI Mascot Interaction System

The AI mascot is a major highlight of the project. It combines the capabilities of DeepSeek V4 Pro and DeepSeek, using carefully designed Prompt Engineering to establish an AI persona centered around the lost and found scenario. The mascot provides three core functions: Deep Analysis (analyzing lost item scenarios), AI Matching (intelligent match recommendations), and Safety Tips (anti-fraud advice).

Users can upload item photos, and the AI combines OCR recognition results for in-depth analysis, returning similar found-item listings.

AI Copywriting Feature

When posting on the forum, users can use the AI copywriting feature. The system intelligently generates copy based on uploaded images, and users can select the length and style. If unsatisfied, they can regenerate; if satisfied, they can apply it with one click.

Community Ecosystem: Forum and Points System

The project also builds a complete community ecosystem. The forum page supports multi-dimensional browsing including All, Trending, and Friends tabs. Posts support comments, likes, bookmarks, and sharing (generating posters). The check-in system rewards 10 vitality points (credits) daily, which can be redeemed for virtual rewards like avatar frames in the redemption center.

On the security front, the project features a three-tier permission architecture: Developer tier (highest privileges, responsible for system maintenance), Admin tier (community moderators who review content and handle reports), and Regular User tier. This is a classic RBAC (Role-Based Access Control) design pattern. In internet community products, permission layering is the foundational infrastructure for ensuring content quality and user safety. The three tiers correspond to different sets of operational permissions: the Developer tier has low-level privileges like direct database operations and system configuration changes; the Admin tier has operational privileges like content review and user management; and the Regular User tier is limited to basic functions like posting, browsing, and messaging. This design follows the Principle of Least Privilege — each role is granted only the minimum set of permissions needed to fulfill its responsibilities, thereby reducing the risk of accidental or malicious operations. Administrators can ban, mute, extend ban durations, or immediately unban users who violate the rules.

Conclusion

As a competition entry, "Campus Find" demonstrates remarkably complete product thinking and technical implementation capabilities. From AI intelligent matching to OCR recognition, from QR code return confirmation to three-tier permission management, the entire system forms a complete closed loop spanning posting, matching, communication, and handoff. The deep integration of DeepSeek's model series transforms this mini program from a mere information posting platform into a truly intelligent AI application with analytical and matching capabilities.

For the campus scenario, a mini program that combines lost and found services, community interaction, and AI capabilities indeed offers significant practical value.