GOAT or TRASH

Home page
Results page
Image Topic Picker Modal

Inspiration

Tier lists are fun, but most tools are manual and static. We wanted to build something that feels like internet culture in real time: type a topic, upload a photo, or rant into the mic, then instantly get a bold GOAT-to-TRASH ranking.
The main idea was to make Gemini 3 the actual product engine, not just a chatbot wrapper.

What it does

GOAT or TRASH is a multimodal tier-ranking app powered by Gemini 3.

Text input: generates a full 5-tier ranking with witty reasons.
Image input: detects whether the image is a menu/list or a scene/object.
Menu/list flow: extracts items and ranks only those items.
Scene/object flow: suggests related ranking topics for user selection.
Voice input: extracts topic + sentiment from audio rant, then builds a personalized ranking.
Share mode: exports the final board as a high-resolution snapshot.

How we built it

We built the app with React + TypeScript (Vite) and integrated @google/genai.

Used gemini-3-flash-preview for low-latency multimodal inference.
Defined strict JSON schemas for structured, predictable output.
Implemented a unified ranking pipeline that supports text, image, and audio contexts.
Added safety checks (isValidTopic + refusal message) for unsafe topics.
Rendered final results as shareable images using html-to-image.
Added robust UI states for loading, errors, and modal-based topic selection.

Challenges we ran into

Keeping responses creative while still returning valid schema-compliant JSON.
Designing reliable image branching (menu/list vs scene/object) across mixed-quality images.
Handling voice-input variability (mic permissions, noisy audio, inconsistent speech).
Balancing response quality and latency for a fast, demo-friendly UX.
Preserving personality while enforcing safety constraints.

Accomplishments that we're proud of

Built a complete end-to-end multimodal experience, not just text generation.
Achieved consistent structured outputs across different input modes.
Created a distinct, shareable product feel with retro UI + one-tap snapshot.
Connected AI reasoning directly to interaction design and user flow.
Delivered a functional, public demo-ready app with clear Gemini-first value.

What we learned

Gemini performs best when we combine strict structure with controlled creativity.
Schema-first design dramatically improves reliability for production-style demos.
Multimodal AI quality depends on product flow as much as model capability.
Good fallback logic and loading/error UX are critical for trust.
Prompt engineering works best when tied to concrete UI states and user intent.

What's next for GOAT or TRASH

Add user accounts, saved lists, and public leaderboard-based voting.
Introduce collaborative party mode for real-time group ranking.
Improve image understanding with richer item normalization and deduplication.
Add multilingual support and regional trend presets.
Build analytics and feedback loops to tune prompts and ranking quality over time.