AI-powered sketch-to-image meta search engine with zero-shot visual discovery.

_{Sketch → visual intent → live internet results (no text input).}

Outlyne uses SigLIP2 and OpenVINO to convert hand-drawn sketches into visual embeddings in real-time (~90ms). It features a zero-shot visual intent layer that interprets sketches to perform live image search across public engines, requiring no text input.

Why Outlyne?

Traditional image search requires text. Outlyne flips the interaction: you sketch first, the system infers intent visually, then searches live.

No indexing. No training. No prompts.

---

Prerequisites

Ensure you have the following installed before starting:

• Python 3.12+
• Bun (JavaScript/TypeScript Task Runner)
• uv (Fast Python package manager)
• Docker (Recommended for production-like environment)

---

Getting Started

1. Clone & Install

git clone https://github.com/itisrohit/Outlyne.git
cd Outlyne

# Install dependencies & setup local cache
bun run sync

2. Verify Vision Core

Run the dedicated benchmark to verify proper model optimization and inference speed:

uv run python tests/bench_embedder.py

3. Start Development

Outlyne features a unified development command that orchestrates both the Python API and the React frontend:

Local Development:

# Start both Backend + Frontend in sync
bun run dev

Docker (Recommended for Backend):

# 1. Build & Launch the API Container
bun run docker:build
bun run docker:up

# 2. Start the Frontend (in a new terminal)
cd web && bun run dev

---

Developer Commands

Command	Description
bun run dev	Unified Dev: Starts API and Web frontend concurrently
bun run sync	Sync Python venv & setup cache dirs
bun run lint	Run Ruff, Mypy, and Biome strict checks across the stack
bun run test	Run the Zero-Shot sketch search verification suite
bun run docker:build	Bake model artifacts into Docker image
bun run docker:up	Spin up the orchestrated stack
bun run clean	Remove all caches, venv, and artifacts

---

Performance Benchmarks

• Visual Encoding: ~92.7ms 🔥 (SigLIP2 on CPU via OpenVINO)
• Semantic Interrogation: ~12ms (Zero-shot intent classification)
• Cold Boot (Docker): ~2s (Vs. 45s locally without pre-baked IR)
• Lints: 100% clean (Strict Mypy + Ruff + Biome)

---

Documentation

• Contributing - Guidelines for contributing to Outlyne.
• Architecture - Deep dive into system design and components.
• Vision Core - Implementation details of the visual engine.