Home Page for authentication
New Registeration page
Sign in page
Home Page after signing in
Filing a lost item
Evidence Page where you try to find your lost item by describing it and seeing if there is a match in the database
If an item has not been collected for over 6 months, you can try your luck to win one by pulling the slot machine!
The dashboard after logging in as an admin.

SherLostHolmes — The Detective for Your Lost Items

Inspiration

We've all been there. Our friend, let's call them "The Forgetful One," had a legendary talent for losing things — keys in the library, a phone in the Hall Building, a jacket somewhere between lectures. Every time, the process was the same: awkward visits to lost-and-found offices, vague descriptions, and the dreaded "we'll call you if something turns up."

We thought: what if reclaiming your lost items felt less like bureaucracy and more like solving a mystery?

That's when SherLostHolmes was born — a gamified, AI-powered lost-and-found system that turns the frustrating experience of losing something into an engaging detective adventure.

What It Does

SherLostHolmes reimagines lost-and-found with:

AI Detective Matching: Using LangChain and vector embeddings, our system semantically matches your lost item description against found items — even if you describe your "blue water bottle" and someone submitted a "navy hydration container"
Blind Lineup: Matched items appear as anonymous "suspects" with blurred images and mysterious personas, preventing false claims
Interrogation Room: An AI persona "guards" each item, asking ownership-verification questions. Only true owners know the secret details!
Fraud Detection: A random decoy item is hidden in every lineup — selecting it instantly flags fraudulent claims
Smart Lockers: ESP32-powered lockers with servo locks allow 24/7 secure pickup with one-time passwords
Lucky Find: A slot machine mini-game where you can win unclaimed items that have been sitting for 6+ months

How We Built It

Backend:

FastAPI (Python) for high-performance async APIs
MongoDB Atlas for flexible document storage
LangChain + OpenRouter for AI orchestration
Custom vector embeddings for semantic search
Cloudinary for image processing (including automatic blurring!)

Frontend:

Next.js 14 with TypeScript
Tailwind CSS for that vintage detective aesthetic
Clerk for authentication
Framer Motion for smooth animations

Hardware:

XIAO ESP32-S3 microcontroller
Servo motor for physical lock mechanism
CardKB I2C keyboard for code entry
RESTful API integration for real-time locker state

Challenges We Faced

Anti-Hallucination in Interrogation: Getting the AI to ask verification questions without accidentally revealing the answers was tricky. We implemented strict prompt engineering with forbidden phrases and validation layers.
Fraud Detection Balance: Adding decoy items to the lineup required careful tuning — too obvious and it's useless, too similar and real owners get confused. We settled on cross-category decoys.
Real-Time Sync: Keeping locker state synchronized between the ESP32 hardware and our backend required careful polling and password refresh mechanisms.

What We Learned

LangChain Agents are powerful but need guardrails — structured workflows beat free-form agent behavior for reliability
Gamification transforms UX — adding detective theming made users want to verify their identity
Hardware + Software integration is incredibly rewarding when it clicks

Built with energy drinks and determination at ConuHacks X