Inspiration
Repeatedly running out of items and forgetting to reorder? Tired of shopping lists and having to go order every single thing you need? Shopi uses Snap Spectacles and real-time computer vision to monitor stock levels and auto reorder the item via voice or gesture.
What it does
- Detects containers with low stock and surfaces replenishment prompts
- Processes images with Gemini via snap-and-buy
- Matches detected items to an on-device catalog
- Supports hand-tracking, voice, and pinch inputs
- Uses AR overlays and TTS for hands-free interaction
- Instantly prompts the reorder of items
How we built it
Frontend (Snap Spectacles):
- Lens Studio with TypeScript/JavaScript
- Spectacles Interaction Kit v0.10.0 for hand gestures
- HandVisual for 3D tracking and visual feedback
- Voice transcription using Snap OS
Backend (FastAPI):
- Python 3+ and AsyncIO
- ngrok for server deployment
- Pydantic for schema validation
- Gemini API (2.5 Flash) for product detection and attribute identification
- Selenium for Shopify checkout automation
- Google Search API for product discovery and label-matching
Challenges we ran into
- Spectacles hand tracking: fixed by enabling disabled HandVisual components in the scene hierarchy
- FastUI integration:
unitPlaneMesherrors were addressed via correct component wiring - Image encoding: large camera frames required base64 JPEG and browser telemetry
- Product matching: refined Gemini prompts to reduce false positives
- Selenium automation: anti-bot mitigations required randomized input timing
Accomplishments that we're proud of
- Working AR glasses integration with real-time vision
- End-to-end path: image → detection → catalog match → purchase
- Hands-free interaction via gestures
- Sub-second AI responses with caching
- Modular FastAPI design enabling future commerce integrations
- Stable build across TS/JS in Lens Studio and FastAPI backends
What we learned
- Cross-platform performance requires hardware-aware optimization
- AR UI benefits from gesture-driven controls and minimal overlays
- Efficient vision-processing pipelines reduce latency
- Error handling and graceful degradation are important
- Fast iteration wins with cloud AI; model fine-tuning remains niche
- Generative AI prompts need specificity to avoid bias and hallucination
What's next for Shopi
- Integrate voice and real-time object detection without need for taking a snapshot
Built With
- brightdata
- fastapi
- gemini
- javascript
- openai
- pydantic
- python
- selenium
- snapos
- spectacles
- typescript

Log in or sign up for Devpost to join the conversation.