Splitshot

Main Interface
Real-Time Object/Human Detection for Bill Splitting
Main Menu for Group Photo Analysis
Computer Vision Algorithm Running
Detected People with Corresponding Items Based on Image
Bill Splitting through Voice Analysis
Share Bill View Option through QR Code
Compact Analysis of Spending Habits with the help of LLM
Roulette Wheel Selection for Bill Paying

Inspiration

Splitting the bill kills the post-meal buzz. For the bunq Hackathon 7.0, we wanted to eliminate the friction of manual math and chasing IBANs. We built Splitshot to make settling the tab as natural as snapping a quick picture or having a conversation.

What it does

AR "Point & Split": Point your camera at the table to see real-time AR tags mapping people, plates, and prices.
Voice Splitting & Disputes: Dictate the order or correct the AI instantly using natural language ("Actually, John had the pasta").
bunq Payments: Settle up instantly via IBAN requests, emails, or shareable bunq.me links.
Smart Dashboard: Track debts and view AI-generated spending summaries in an Instagram-style interface.

How we built it

Frontend: Next.js 14, Tailwind CSS, Framer Motion.
AI Engine: Claude Sonnet (vision/AR tracking), Claude Haiku (text parsing), and Deepgram Nova-3 (audio transcription).
Backend & Storage: PostgreSQL (Neon), Drizzle ORM, AWS S3 (face enrollments).
Banking: bunq Sandbox REST API with strict RSA-signed authentication.

Challenges we ran into

Browser AR Latency: Keeping live camera overlays smooth despite API delays required building custom interpolation animations.
Multimodal Merging: Combining fuzzy voice descriptions, visual table analysis, and receipt OCR into a single, accurate JSON output.
Cryptographic Security: Implementing bunq's complex 3-step RSA-signed authentication within a serverless architecture.

Accomplishments that we're proud of

"Magical" Voice Disputes: Allowing users to seamlessly correct visual AI mistakes simply by talking to the app.
Live Spatial Reasoning: Pushing Claude Vision to map 2D images to physical table dynamics in real-time.
Hackathon Polish: Delivering a beautifully animated, fully functional full-stack banking app under strict time constraints.

What we learned

How to maximize LLM spatial reasoning capabilities for real-world physical mapping.
The intricacies of securely integrating European banking APIs.
That giving users an intuitive audio fallback to correct visual AI drastically improves trust in the product.

What's next for Splitshot

Go Live: Migrating from the sandbox to the production bunq environment for real-money requests.
Native Apps: Building dedicated iOS and Android apps to unlock hardware-accelerated AR tracking.
Trip Splitting: Expanding to handle multi-day, multi-currency group vacations with algorithmic debt simplification.