Pic2cook

Inspiration

Traditional recipe services are trapped in the "Static Search" paradigm. They force users to navigate a high-friction entry barrier: remembering exact ingredient names or struggling to articulate the contents of a diverse refrigerator. Pic2Cook was born from a simple realization: Visual context is the shortest path to action. We transitioned from a "lookup model" to a "vision-first partner," transforming visual ambiguity into precise, delicious culinary outcomes.

What it does

Pic2Cook is an "action-oriented" culinary ecosystem designed to own the entire journey from vision to table:

"Picture-to-Recipe" Core: Capture any dish or ingredient set to generate an instant, comprehensive recipe.
Infinite Personalization: Unlike static databases, our generated recipes are dynamic templates. Users can infinitely customize portions, swap ingredients (via AI-suggested alternatives), and modify techniques to fit their specific needs.
Multilingual & Inclusive Intelligence: Native support for English, Korean, and Japanese. The system provides real-time allergy alerts, dietary filtering (Vegan, Keto, etc.), and automated nutritional breakdowns.
Hyper-Local Mapping: Bridge the gap between "Global Taste" and "Local Ingredients." The AI identifies hard-to-find components and suggests accessible local substitutes based on the user's current geography.
Interest-Based Social Network: A trending feed where users follow creators (e.g., Culinary Class Wars influencers) and get real-time culinary updates.

How we built it (AI-Native Stack)

We utilized a cutting-edge AI-Native development workflow to achieve high-speed iterations and architectural reliability:

Intelligence Engine: Gemini 3 (integrated via the Pic2Cook API) provides the backbone for multi-modal reasoning, enabling complex image analysis and nutritional inference.
Agentic Orchestration (Antigravity & oh-my-ag): We didn't build this project alone. We used Antigravity, an agentic coding partner, and our proprietary oh-my-ag orchestrator to manage role-based sub-agents (PM, Backend, Frontend, QA). This ensured clean, synchronized state across FastAPI, Next.js, and Flutter.
Cloud Infrastructure (GCP): Powered by Google Cloud Platform, utilizing Google Vision API for semantic analysis, Cloud Run/Tasks for asynchronous processing, and Vertex AI for model management.

Challenges & Lean Validation

Defeating "Hallucinations" with Real-World Feedback: Skeptical of standard LLM outputs, our team personally cooked and validated AI-generated recipes over 50 times. Each failed meal was used to iteratively refine our prompt engineering and logic, ensuring professional-grade reliability.
Performance vs. Latency: Complex AI reasoning can be slow. We solved the UX gap by implementing Server-Sent Events (SSE), providing users with a "transparent thinking" phase that visualizes progress in real-time.

Accomplishments that we're proud of

In addition to the Pic2Cook service, we have contributed to the AI-Native developer ecosystem:

Fullstack Starter: A battle-tested template for building AI-Native, multi-platform applications.
oh-my-ag: A multi-agent orchestrator that mitigates model volatility by distributing responsibilities across specialized AI roles.

What we learned

The project proved that in the AI era, context is the new currency. A model's sheer power is secondary to how it integrates with real-world constraints (like what’s actually in your fridge). We also learned that moving from "Single-Prompt" to "Multi-Agent" orchestration is the only way to build reliable, high-complexity systems.

What's Next: The Hyper-Local Evolution

"Grandmother’s Secret" (OCR Logic): We are finalizing a feature to digitize analog family secrets—transforming faded, cursive handwritten notes into structured digital recipes via Gemini’s advanced OCR.
The Culinary Hub: Expanding into direct-to-retail integrations, allowing users to order missing ingredients from local markets with a single tap, effectively bridging the gap between digital inspiration and physical ingredients.